v0.1.903-beta

This release adds a browser inside Unsloth (browser use coming very soon), so files, web pages and pages the model writes open right beside your chat. It also brings [EmbeddingGemma 2](https://unsloth.ai/docs/models/embeddinggemma-2), Google's new multimodal embedding model, plus new Audio pages and lower-memory training.

## Highlights

- **Browser beside chat** for files, web pages and model-written HTML, as tabs
- **EmbeddingGemma 2**: run and fine-tune Google's new model for text, code, images, video and audio [Guide](https://unsloth.ai/docs/models/embeddinggemma-2)
- **Audio pages**: Speak, Clone, Music, Transcribe, Edit, Separate, Convert, with 70+ GGUF models
- **Keep multiple models loaded**, plus vLLM and SGLang engines (opt-in)
- **Sandbox** for code the model runs, now with a Settings tab and Colab and Docker support
- **Live preview** during image and video generation

## Browser

Unsloth now has a built-in browser panel next to any chat. Attached files, web pages and HTML the model writes all open in it as tabs.

- **Files open as tabs**: PDF, Office, HTML, Markdown, code and text.
- **Real browsing**: address bar, search, tabs, find in page, zoom, bookmarks, history and downloads.
- **Model-written HTML and code open as tabs** with a preview, console and Request edits, replacing Canvas.
- **Ask about this page**: highlight part of a page and ask the model about it. Right-click a tab to pin it or fork it into a chat.
- **Settings > Browser**: pick a search engine, set the default zoom, and import bookmarks from Chrome, Safari, Firefox or Edge.
- **Kept separate**: web pages cannot reach your Unsloth account, chats or API.

## EmbeddingGemma 2

EmbeddingGemma 2 is Google DeepMind's new multimodal embedding model. It turns text, code, images, video and audio into vectors for search, RAG, classification and clustering. Read our [guide](https://unsloth.ai/docs/models/embeddinggemma-2).

- **740M parameters, or 270M for text only**, with optional vision and audio encoders. 100+ languages and an 8K context window.
- **About 14% better on code** than EmbeddingGemma 1 (78.68 vs 68.76, MTEB code).
- **Run it locally**: import GGUF or safetensors into Unsloth, or serve text embeddings with llama.cpp. [GGUF](https://huggingface.co/unsloth/embeddinggemma-2-GGUF) [safetensors](https://huggingface.co/unsloth/embeddinggemma-2)
- **Fine-tune the text model** with `FastSentenceTransformer`, using 4-bit QLoRA or BF16 LoRA.
- **Prompts are kept.** Fine-tunes of EmbeddingGemma and Qwen3-Embedding now save the model's built-in query and document prompts, not empty ones.

## Audio

- **Clone** a voice from a short sample and save it.
- **Edit** words in a recording and keep the speaker's voice.
- **Separate** a song into vocals, drums and other parts, or **Convert** one voice into another.
- **Music** makes songs and sound effects, and can edit parts of a clip.
- **Transcribe** can show timestamps and speakers, and exports subtitles.

## Sandbox

Code the model runs with the Python and Terminal tools stays inside a sandbox, with limited access to your computer.

- **OS sandbox** on Linux and macOS, plus an opt-in Windows 11 preview. Now works in Colab and Docker.
- **Settings > Sandbox** shows this computer's protection and can install a missing sandbox.
- **Low or High**: High (default) uses the OS sandbox. Low uses software checks only.

## Models + chat

- **Keep multiple models loaded** (off by default, in Settings > Resources). Each request is answered by the model it asks for.
- **vLLM and SGLang**: opt-in, experimental engines for Linux with NVIDIA GPUs.
- **Auto-compact** now works for API models like Claude and OpenAI, and for MLX chats.
- Import your Cursor, Claude Code and Codex chats into projects, or chats saved as Markdown.
- `unsloth start --app` lets the Codex, OpenCode, OpenClaw and Hermes desktop apps use your local model.

## Training + speed

- **Layer offload** (opt-in, LoRA, NVIDIA and AMD graphics cards): `offload_layers` keeps frozen layers in system RAM, so Llama-3.1-8B 4-bit trains with VRAM capped at 7 GiB.
- **Faster padding-free training without flash-attn**, the usual Colab case: Qwen3-0.6B drops from 5.45 GB to 1.87 GB of VRAM, with steps 5x to 8x faster.
- **gpt-oss-20b QLoRA is 2x to 3x faster** on NVIDIA.
- **Images and video** (NVIDIA): live preview while rendering, up to 20x faster Z-Image-Turbo on older GPUs, and the same seed gives the same image after a restart.
- **AMD**: export fine-tuned models for the Ryzen AI NPU, and a much faster first Qwen-Image-2.1 image on newer Radeon cards.
- **Faster on Apple Silicon (MLX)**: Qwen3.6-35B-A3B replies about 10% faster, and long prompts start up to 10% sooner.
- **Int8 Prefill on M5 chips** (MLX, experimental, opt-in): long prompts read up to 1.5x faster, with slightly less accurate answers.

## Download Unsloth Desktop

Unsloth Desktop is **free and open source**. Download it for:
<table>
  <tr>
    <td><b>Platform</b></td>
    <td><b>Link</b></td>
  </tr>
  <tr>
    <td><b>Windows</b></td>
    <td><a href='https://github.com/unslothai/unsloth/releases/latest/download/Unsloth-Desktop-Windows.exe'>Download</a></td>
  </tr>
  <tr>
    <td><b>macOS</b></td>
    <td><a href='https://github.com/unslothai/unsloth/releases/latest/download/Unsloth-Desktop-MacOS.dmg'>Download</a></td>
  </tr>
  <tr>
    <td><b>Linux x64 / Ubuntu (deb)</b></td>
    <td><a href='https://github.com/unslothai/unsloth/releases/latest/download/Unsloth-Desktop-Ubuntu.deb'>Download</a></td>
  </tr>
  <tr>
    <td><b>Linux ARM64 / Ubuntu 24.04+ (deb)</b></td>
    <td><a href='https://github.com/unslothai/unsloth/releases/latest/download/Unsloth-Desktop-Ubuntu-ARM64.deb'>Download</a></td>
  </tr>
  <tr>
    <td><b>Linux x64 (AppImage)</b></td>
    <td><a href='https://github.com/unslothai/unsloth/releases/latest/download/Unsloth-Desktop-Linux.AppImage'>Download</a></td>
  </tr>
  <tr>
    <td><b>Windows ARM64</b></td>
    <td><a href='https://github.com/unslothai/unsloth/releases/latest/Unsloth-Desktop-Windows-ARM64.exe'>Download</a></td>
  </tr>
</table>

## What's Changed
* feat(studio): complete embedded image recipe and Recipe popover by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/12091
* Bump install.sh / install.ps1 pin to unsloth>=2026.9.14 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12439
* Count the final EOS token within the raw-text chunk budget by @vineethsaivs in https://github.com/unslothai/unsloth/pull/11040
* fix(studio): serve desktop SPA on loopback listener by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10847
* Format the desktop contract test to the ruff-format fixed point by @danielhanchen in https://github.com/unslothai/unsloth/pull/12440
* Fix three tests red on main after #12431, #12408 and #12382 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12441
* Studio: disclose required assets before media downloads by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12387
* Stop enable_padding_free_metadata writing seq_lengths into the caller's examples by @vineethsaivs in https://github.com/unslothai/unsloth/pull/8049
* Studio: persist Data Recipes and run history server-side by @Etherll in https://github.com/unslothai/unsloth/pull/7108
* Fix the tests left red on main by #12408, #12382 and #12436 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12447
* Read supervision from labels or the mask column in the padding-free filter tests by @danielhanchen in https://github.com/unslothai/unsloth/pull/12452
* fix(studio): include loaded llama extra args in active model baseline by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10870
* Read auto_map at every config level in the remote code gate by @danielhanchen in https://github.com/unslothai/unsloth/pull/12453
* Record why the permission pill is missing after a reload, and reload once more only if the app never booted by @danielhanchen in https://github.com/unslothai/unsloth/pull/12456
* Add approved image attachment inputs for MCP tools by @Etherll in https://github.com/unslothai/unsloth/pull/10871
* Studio: log a pending embedder download once instead of a traceback on every compaction by @danielhanchen in https://github.com/unslothai/unsloth/pull/12443
* Studio: stop a recovered chat crashing on a duplicate tool call key by @danielhanchen in https://github.com/unslothai/unsloth/pull/12442
* Give the NVIDIA probe control case a budget a loaded runner can meet by @danielhanchen in https://github.com/unslothai/unsloth/pull/12462
* Wait for the chat-only export gate instead of counting it as the form appears by @danielhanchen in https://github.com/unslothai/unsloth/pull/12463
* Pin the DeepSeek OCR module fetch and import it from that fetch by @danielhanchen in https://github.com/unslothai/unsloth/pull/12454
* Check a sentence-transformers module config on every version, and confirm the resolved class by @danielhanchen in https://github.com/unslothai/unsloth/pull/12444
* Studio: keep the app's titlebar chrome off the desktop update screen by @oobabooga in https://github.com/unslothai/unsloth/pull/12465
* Studio: stop a reopened chat from saving every replayed event, and let Stop reach it by @oobabooga in https://github.com/unslothai/unsloth/pull/12464
* Studio: blur modal titlebar background while keeping window controls sharp by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12433
* Gate the sentence-transformers module types on the routes that delegate the load by @danielhanchen in https://github.com/unslothai/unsloth/pull/12446
* Retry the NVIDIA probe control case past the Linux pwsh redirect race by @danielhanchen in https://github.com/unslothai/unsloth/pull/12477
* Studio: keep diffusion LoRA loading working on torchao 0.18 with peft 0.18 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12494
* Studio: retry Lemonade on a new port when its free port is taken before it binds by @danielhanchen in https://github.com/unslothai/unsloth/pull/12498
* Studio: move Butterfly Pea and Earl Grey up the More themes list, and rename Cherry Cola to Cherry by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12500
* fix(studio): prevent IME confirmation from submitting chat renames by @kei-yamazaki in https://github.com/unslothai/unsloth/pull/12475
* Wait for vite to serve before the titlebar, find, shortcuts and settings smokes navigate by @danielhanchen in https://github.com/unslothai/unsloth/pull/12501
* Studio: top-level Models block group in picker by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10736
* fix(studio): keep distinct symlink aliases for per-model settings by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10644
* Studio: JSON and Markdown validator blocks by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10710
* fix(studio): reconcile externally updated saved assistant messages by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/11500
* Studio: save a merged Whisper export that can be loaded again by @NilayYadav in https://github.com/unslothai/unsloth/pull/12483
* fix(studio): allow Enter to send with idle macOS Pinyin IME by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/12138
* Studio: keep a Data Recipes publish private when the Hub repo already exists by @NilayYadav in https://github.com/unslothai/unsloth/pull/12482
* Baseline the nine unsloth-zoo 2026.9.9 findings after review by @danielhanchen in https://github.com/unslothai/unsloth/pull/12515
* Studio: put llama-server warnings and errors in the default session log by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10848
* Studio: stop leftover dataset columns becoming the system prompt by @NilayYadav in https://github.com/unslothai/unsloth/pull/12485
* Studio: restore markdown chat import and bulk export by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/12513
* Add install_missing_dependencies to opt out of the FP8/FP4 llm-compressor auto-install by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/9405
* perf(studio): enter the MLX routed-experts and norm-handoff fusions per request by @Lyxot in https://github.com/unslothai/unsloth/pull/12422
* Feat/UI rework by @Sneakr in https://github.com/unslothai/unsloth/pull/12479
* Studio: run MiniMax-H3's Diffusers INT8 path on 24 / 16 / 12 GB cards and near the host RAM floor by @danielhanchen in https://github.com/unslothai/unsloth/pull/12409
* Studio: extend the ROCm MIOpen cutoff and Qwen-Image-2.1 query chunking beyond gfx1151 by @oobabooga in https://github.com/unslothai/unsloth/pull/12471
* fix(studio): block-split long backslash lines without inline lex (#11376) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/11501
* Studio: train vision models on images in a list column or inside the messages by @NilayYadav in https://github.com/unslothai/unsloth/pull/12484
* Studio: resume a reply that stopped mid-thought on MLX by @Lyxot in https://github.com/unslothai/unsloth/pull/12504
* Studio: canonicalize and validate web_search arguments by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/9716
* Studio: optional systemd user service for Linux auto-start by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/9308
* SentenceTransformer: opt-in encoder unpadding via shared attention dispatch by @Etherll in https://github.com/unslothai/unsloth/pull/4460
* Studio: live dataset preview during active runs by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10737
* Core: pip extras and auto-install support for torch 2.13 and 2.14 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12151
* Studio: resume NPU model downloads, keep their progress, and list NPU models in the Hub by @oobabooga in https://github.com/unslothai/unsloth/pull/12512
* Name the matching build when flash-attn or causal-conv1d was built for another torch by @danielhanchen in https://github.com/unslothai/unsloth/pull/12211
* Studio: fused int8 GEMM with the dequant epilogue, bf16 out (Qwen-Image-2.1 14-17% faster per step) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12448
* Studio: wait for the torch warm's dynamo import before a load's first download by @danielhanchen in https://github.com/unslothai/unsloth/pull/12472
* Studio: faster cold start for image and video loads (first image on H100 7 to 37% sooner) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12496
* Studio: let the compile cache keep denoiser graphs with sdpa_kernel blocks (restart first render 1-5.5 s faster) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12510
* Studio: keep GGUF weights and the last-run module on their host copy under whole-model offload by @danielhanchen in https://github.com/unslothai/unsloth/pull/12022
* Studio: torch 2.13 for new Linux cu130 Python 3.13 installs, existing installs keep their torch by @danielhanchen in https://github.com/unslothai/unsloth/pull/12150
* Studio: plan image offload on measured activations so more VRAM is never slower by @danielhanchen in https://github.com/unslothai/unsloth/pull/12043
* studio: reuse ssl setup for local gguf requests by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12495
* Studio: run FLUX.2 RoPE as one fused kernel on fp16 GPUs (klein step 5% faster on T4, bit-identical) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12503
* Studio: load FLUX.1 and Qwen-Image on fp16-only GPUs with small host RAM by @danielhanchen in https://github.com/unslothai/unsloth/pull/12480
* Studio: run Z-Image and Wan2.2-5B / HunyuanVideo-1.5 in fp16 on fp16-only GPUs instead of fp32 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12476
* Studio: faster first renders and resident text encoder under group offload by @danielhanchen in https://github.com/unslothai/unsloth/pull/12519
* Studio: make unsloth start sampling and reasoning flags apply only to that agent by @NilayYadav in https://github.com/unslothai/unsloth/pull/12493
* Studio: stop adding backslashes to Word files in Data Recipes by @NilayYadav in https://github.com/unslothai/unsloth/pull/12488
* Studio: keep tool-call arguments when training on tool-calling datasets by @NilayYadav in https://github.com/unslothai/unsloth/pull/12487
* Studio: keep <placeholder> and Vec<T> text in research reports and file previews by @NilayYadav in https://github.com/unslothai/unsloth/pull/12489
* Studio: keep Word equations when a .docx is added to a knowledge base or chat documents by @NilayYadav in https://github.com/unslothai/unsloth/pull/12490
* Studio: keep MiniMax-H3 GGUF renders on a resident sd-server, released when idle by @danielhanchen in https://github.com/unslothai/unsloth/pull/12461
* Studio: load the cached int8 checkpoint for a 5-bit-or-below GGUF pick that has to offload (Qwen-Image-2.1 about 2x faster at 12 / 8 GB) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12455
* Studio: log the real reason the Decision API model failed to load by @NilayYadav in https://github.com/unslothai/unsloth/pull/12491
* Studio: stop CLI commands running as another account when Studio has more than one by @NilayYadav in https://github.com/unslothai/unsloth/pull/12492
* Studio: train a dataset's system column as the system prompt by @NilayYadav in https://github.com/unslothai/unsloth/pull/12486
* Studio: add per-model custom llama.cpp INI configuration by @Etherll in https://github.com/unslothai/unsloth/pull/10783
* Fused Triton NF4 dequant and GEMV: bit-exact, stream-safe, torch.compile traceable by @danielhanchen in https://github.com/unslothai/unsloth/pull/12113
* Make RMSNorm, RoPE and the input-embedding hook traceable by torch.compile by @danielhanchen in https://github.com/unslothai/unsloth/pull/12171
* Studio: add opt-in vLLM and SGLang support with multi-GPU inference, quantization and vision by @oobabooga in https://github.com/unslothai/unsloth/pull/11491
* Lift the transformers ceiling to 5.17.0 and align the mirrored CI caps by @danielhanchen in https://github.com/unslothai/unsloth/pull/11016
* Studio: run vLLM and SGLang on Windows inside a private WSL2 distro by @danielhanchen in https://github.com/unslothai/unsloth/pull/12024
* Studio: add audio.cpp as a native engine for speech, music and dictation by @Etherll in https://github.com/unslothai/unsloth/pull/12342
* Studio: draw the command palette like chat search by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12511
* Studio: center the sidebar row kebab and pin in their hover circle by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12502
* Studio: clip every rounded scroller in Firefox, dropping the overflow watcher by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12438
* Studio: keep the composer permission shield still while the plus spins by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12506
* Parity: keep the reap-margin check off the edge of its own window by @lonexreb in https://github.com/unslothai/unsloth/pull/12514
* Studio: use ComfyUI's default settings for FLUX.1, Qwen-Image, Z-Image, Ideogram 4, Wan2.2 and HunyuanVideo-1.5 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12516
* Studio: move torchao int8 weights back to the GPU after an oversized request streams pinned groups by @danielhanchen in https://github.com/unslothai/unsloth/pull/12526
* Count only the lock's own waits in the unlockable-filesystem row by @lonexreb in https://github.com/unslothai/unsloth/pull/12529
* Studio: renew the chat-run lease while long prefill is still advancing by @indrajeetapache in https://github.com/unslothai/unsloth/pull/12523
* Studio: stop rejecting --mmproj-device CUDA1 when gpu_ids are saved by @1fanwang in https://github.com/unslothai/unsloth/pull/12505
* fix(studio): WSL2 Windows localhost hint in startup banner (#11187) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/11361
* Studio: lift a carried model picker row like a sidebar chat, and lighten both drag copies by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12499
* Postpone annotation evaluation in engine_compat.py so main's PEP 604 ratchet holds by @danielhanchen in https://github.com/unslothai/unsloth/pull/12556
* Keep the MLX inference test stub neutral for every helper zoo enters around a model by @danielhanchen in https://github.com/unslothai/unsloth/pull/12548
* Re-measure the Studio startup budget after the managed engine options by @danielhanchen in https://github.com/unslothai/unsloth/pull/12560
* Studio: nudge sidebar row pin and options buttons 3px right by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12563
* Studio: make carried rows a faintly frosted, darker copy by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12562
* Check the sidebar pin follows the UI scale by property, not by its pinned length by @danielhanchen in https://github.com/unslothai/unsloth/pull/12597
* Studio: search chats, projects, files and models from tabs in chat search by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12564
* Revert 10 PRs merged before Codex review converged by @Etherll in https://github.com/unslothai/unsloth/pull/12569
* Studio: cleaner Deep research limits dialog by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12632
* Studio: Skills dialog scrolls on its right edge by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12634
* Studio: Skills pill in the composer for quick toggling by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12631
* Studio: shorter permission menu, Learn more, calmer Full access by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12630
* Studio tests: pin Z-Image's fp16 promotion in the calibrated-activation test by @danielhanchen in https://github.com/unslothai/unsloth/pull/12570
* Studio: keep text placed over pictures when indexing a PDF by @NilayYadav in https://github.com/unslothai/unsloth/pull/12577
* Parity: keep the reap-margin check off the edge of its own window by @danielhanchen in https://github.com/unslothai/unsloth/pull/12615
* Count only the lock's own waits in the unlockable-filesystem row by @danielhanchen in https://github.com/unslothai/unsloth/pull/12618
* Studio: stop blaming Max Tokens when a connected model fills its context window by @oobabooga in https://github.com/unslothai/unsloth/pull/12567
* Studio: keep a tool call's answer after its result in JSONL chat exports by @NilayYadav in https://github.com/unslothai/unsloth/pull/12579
* Studio: stop unsloth start openclaw cutting every reply at 8,192 tokens by @NilayYadav in https://github.com/unslothai/unsloth/pull/12573
* Studio: stop New chat from requesting a thread row that does not exist yet by @oobabooga in https://github.com/unslothai/unsloth/pull/12539
* Studio: stop training Gemma 4 on tool results by @NilayYadav in https://github.com/unslothai/unsloth/pull/12575
* Studio: say so when the Audio page stops speech at Max tokens by @NilayYadav in https://github.com/unslothai/unsloth/pull/12583
* Studio: lift a carried model picker row like a sidebar chat, and lighten both drag copies by @danielhanchen in https://github.com/unslothai/unsloth/pull/12622
* Studio: show NPU reply speeds and configure NPU models before loading by @oobabooga in https://github.com/unslothai/unsloth/pull/12595
* Studio: consent dialog before FP8/FP4 llm-compressor install (Phase 2) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/9554
* Studio: train a dataset's system column as the system prompt by @danielhanchen in https://github.com/unslothai/unsloth/pull/12614
* Studio: stop rejecting --mmproj-device CUDA1 when gpu_ids are saved by @danielhanchen in https://github.com/unslothai/unsloth/pull/12620
* fix(studio): prevent IME confirmation from submitting chat renames by @danielhanchen in https://github.com/unslothai/unsloth/pull/12613
* Chat UI: find the Full access consent dialog by its slots, not its wording by @danielhanchen in https://github.com/unslothai/unsloth/pull/12640
* Studio: read text files saved in older Windows encodings correctly by @NilayYadav in https://github.com/unslothai/unsloth/pull/12581
* Studio: use the reply on screen when turning chats into training data by @NilayYadav in https://github.com/unslothai/unsloth/pull/12584
* fix(studio): WSL2 Windows localhost hint in startup banner (#11187) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12621
* Studio: let API requests that ask for a JSON reply still call their tools by @NilayYadav in https://github.com/unslothai/unsloth/pull/12578
* Studio: keep the VRAM figure visible on the Run preview hardware row by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12545
* Unsloth Studio / Desktop: keep release-notes table links from splitting mid-word in the update popup by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12543
* Studio: keep the prompt queue's more menu open while queued prompts send by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12546
* Let agent desktop apps use the model Unsloth is serving by @NilayYadav in https://github.com/unslothai/unsloth/pull/12582
* Studio: MiniMax-H3 GGUF uses the BF16 cuBLAS matmul path by default on sm80+ (1.2 to 2.3x per step, closer to an F32 reference) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12527
* Studio: renew the chat-run lease while long prefill is still advancing by @danielhanchen in https://github.com/unslothai/unsloth/pull/12619
* Studio: use ComfyUI's default settings for FLUX.1, Qwen-Image, Z-Image, Ideogram 4, Wan2.2 and HunyuanVideo-1.5 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12616
* Studio: keep earlier tool calls in chat history on safetensors and MLX models by @NilayYadav in https://github.com/unslothai/unsloth/pull/12574
* Unsloth Studio / Desktop: stop spam-clicking sidebar rows from queueing a navigation per click by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12544
* Studio: stop the Load Model panel flagging Auto context loads as over VRAM by @oobabooga in https://github.com/unslothai/unsloth/pull/12555
* Windows ROCm: use math attention where the fused SDPA kernels fail by @danielhanchen in https://github.com/unslothai/unsloth/pull/12639
* Studio: stream a torchao denoiser from an unpinned host copy instead of leaving it on the GPU by @oobabooga in https://github.com/unslothai/unsloth/pull/12551
* Point five mapper rows at their upstream repo instead of an unpublished unsloth 16bit name by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12587
* Studio small-host route: prefetch the streamed text encoder and fuse the int8 dequant (T4 FLUX.1-schnell 5.19 to 4.49 s per step, pixel-identical) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12532
* studio: allow full access for the installation owner by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12604
* Keep the load-time gradient checkpointing mode in for_training by @danielhanchen in https://github.com/unslothai/unsloth/pull/12629
* Studio: move torchao int8 weights back to the GPU after an oversized request streams pinned groups by @danielhanchen in https://github.com/unslothai/unsloth/pull/12617
* Studio: decode a resident Wan2.2-TI2V-5B VAE untiled when it fits by @danielhanchen in https://github.com/unslothai/unsloth/pull/12603
* Studio: keep a load's diffusers / peft import off the post-warm import by @danielhanchen in https://github.com/unslothai/unsloth/pull/12572
* Studio: pin inductor's dynamic_scale_rblock off so compiled renders match across servers by @danielhanchen in https://github.com/unslothai/unsloth/pull/12588
* Studio int8 GEMM: stop the device probe from reserving 0.7 to 1.4 GB of spill memory by @danielhanchen in https://github.com/unslothai/unsloth/pull/12525
* Studio: keep dark ink visible in transparent source images by @NilayYadav in https://github.com/unslothai/unsloth/pull/12576
* Studio: make the fused int8 kernels bit-exact on torchao 0.17 and round int32 to bf16 twice like PyTorch by @danielhanchen in https://github.com/unslothai/unsloth/pull/12633
* Add a fast lint gate for risky loader call sites by @danielhanchen in https://github.com/unslothai/unsloth/pull/12636
* Studio: fused fp16 Wan block kernels on T4 (Wan2.2-5B 11.7 to 10.0 s per step, within base's spread) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12531
* Studio: serve multiple models at once by @NilayYadav in https://github.com/unslothai/unsloth/pull/11591
* Fix text-only VLM decoder saves reloading with random weights by @1fanwang in https://github.com/unslothai/unsloth/pull/12561
* Studio: hold cudnn.benchmark off for Wan so renders are reproducible across servers by @danielhanchen in https://github.com/unslothai/unsloth/pull/12565
* Studio: hold cudnn.benchmark off for Qwen-Image, HunyuanVideo-1.5, FLUX.1, Z-Image and SDXL by @danielhanchen in https://github.com/unslothai/unsloth/pull/12568
* Studio: MiniMax-H3 attention fast path and fused q/k norm + RoPE (5-7% faster per step) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12459
* Studio: MiniMax-H3 int8 Linears take the fused-dequant int8 GEMM (A100 9% faster per step, ahead of ComfyUI) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12507
* Studio: fuse MiniMax-H3's ConvRot rotation into the int8 activation quant by @danielhanchen in https://github.com/unslothai/unsloth/pull/12517
* Keep torch deprecation warnings at the caller through the __getattr__ wrapper by @danielhanchen in https://github.com/unslothai/unsloth/pull/12635
* Studio: let small models read a long attached file when asked to summarize it by @NilayYadav in https://github.com/unslothai/unsloth/pull/12580
* Studio: engage the fused int8 GEMM when the placement pins the whole denoiser, and restore torchao weights after an oversized request by @danielhanchen in https://github.com/unslothai/unsloth/pull/12524
* studio: let vision models view workspace images in code mode by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12591
* Studio: keep the whole int8 Qwen-Image-2.1 denoiser resident at 12 GB, released only while the encoders run (298 to 212 ms per step) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12536
* Record the Wan fused block's import_module in the risky loader baseline by @danielhanchen in https://github.com/unslothai/unsloth/pull/12643
* Studio: stop Whisper dropping sentences from clips longer than 30 seconds by @NilayYadav in https://github.com/unslothai/unsloth/pull/12481
* Accept a list train_dataset again on TRL 1.10+ (fixes vision notebooks) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12628
* Unsloth Studio: export or convert a GGUF to FastFlowLM Q4NX for the AMD Ryzen AI NPU by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12541
* Studio: prefetch an image load's weights and upload host tensors through a pinned ring by @danielhanchen in https://github.com/unslothai/unsloth/pull/12590
* Studio: rounder, roomier toasts by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12655
* Follow the multi-model refactor in two guards that went red on main by @danielhanchen in https://github.com/unslothai/unsloth/pull/12653
* Harden torch.export .pt2 loading against CVE-2026-4538 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12657
* Guard against transformers config and chat template CVEs on older versions by @danielhanchen in https://github.com/unslothai/unsloth/pull/12658
* Studio: bump PyJWT, urllib3, Pillow and frontend overrides for security advisories by @danielhanchen in https://github.com/unslothai/unsloth/pull/12656
* Studio: gate repo-hosted embedder modules, MLX model_file, and _socket by @danielhanchen in https://github.com/unslothai/unsloth/pull/12659
* Harden npm and yarn installs against supply-chain attacks by @danielhanchen in https://github.com/unslothai/unsloth/pull/12660
* Studio: encode video mp4 with x264 frame threads (export 1.25 to 2.5x faster, same settings) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12648
* Studio: keep managed accounts off owner files and device nodes by @danielhanchen in https://github.com/unslothai/unsloth/pull/12661
* Studio: stop the Windows event loop spinning when its self-pipe is closed by @Etherll in https://github.com/unslothai/unsloth/pull/12665
* Studio: skip the unused Qwen-Image-2.1 text-encoder lm_head (0.6 to 1.3 GiB less VRAM, bit-identical images) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12669
* Unsloth Studio: managed accounts keep their model settings and pins across an account switch by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12549
* Keep embedding optimizer state 32-bit when embedding_learning_rate is set by @vineethsaivs in https://github.com/unslothai/unsloth/pull/12458
* Studio: never fail, slow down or render noise on an explicit SageAttention or FlashAttention 4 request by @danielhanchen in https://github.com/unslothai/unsloth/pull/12641
* Studio: SageAttention 2 from the kernels hub, and FlashAttention 4 dependencies that load on a fresh install by @danielhanchen in https://github.com/unslothai/unsloth/pull/12654
* Studio: load the hosted FP8 checkpoints for FLUX.1-dev, FLUX.2-klein-4B, FLUX.2-dev and Qwen-Image-2512, and seed Krea 2 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12649
* Studio: stream offloaded denoiser blocks without host waits (B200 FLUX.1 16 GB 0.59x, L4 0.90x per step) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12664
* Studio: pin inductor's reduction configs for LTX-2 compiles so renders match across servers by @danielhanchen in https://github.com/unslothai/unsloth/pull/12651
* Studio: set the ROCm AOTriton opt-in from the inference package so every entry point gets fused attention by @danielhanchen in https://github.com/unslothai/unsloth/pull/12668
* Studio frontend: bump hono, ip-address, qs, dompurify and @babel/core for security advisories by @danielhanchen in https://github.com/unslothai/unsloth/pull/12679
* Studio: clarify image preview loading and retry failed downloads by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12450
* Studio (AMD, Windows): pin the multi-arch ROCm torch to rocm7.14.0, whose fused attention works by @danielhanchen in https://github.com/unslothai/unsloth/pull/12670
* fix: preserve activation QAT in LoRA MLPs by @taking-lying-flat in https://github.com/unslothai/unsloth/pull/12558
* Installers: resolve studio from the venv, never from the caller's directory by @danielhanchen in https://github.com/unslothai/unsloth/pull/12682
* Studio: read pre-quantized diffusion checkpoints from safetensors on torchao 0.17, 0.18 and main by @danielhanchen in https://github.com/unslothai/unsloth/pull/12645
* Studio: give the MiniMax-H3 sd.cpp engine a CUDA-12 cuDNN for fused attention on sm80+ Linux by @danielhanchen in https://github.com/unslothai/unsloth/pull/12662
* Apply the DoRA magnitude in fast_linear_forward by @vineethsaivs in https://github.com/unslothai/unsloth/pull/12642
* Studio: faster first start for image loads (VAE kernel prebuild, early quant probe, one FLUX.1 single-block graph) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12666
* Studio: do not reserve the OS share twice on Linux ROCm APUs by @danielhanchen in https://github.com/unslothai/unsloth/pull/12675
* Studio: honour an explicit flash attention request on ROCm (gfx1151 8 to 12% faster per step) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12667
* Studio projects: folder-plus icon for sources by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12685
* Studio: channels_last for 3D-conv VAEs and untiled A14B decode by @danielhanchen in https://github.com/unslothai/unsloth/pull/12672
* [XPU] Enable FP8 training on XPU by @JoshuaL3000 in https://github.com/unslothai/unsloth/pull/12535
* Studio: per-model automatic step skip at each model's measured step count (1.44 to 1.81x, MiniMax-H3 on max) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12652
* Studio: deterministically load explicitly mentioned skills before generation by @wasimysaid in https://github.com/unslothai/unsloth/pull/12538
* Studio: run fp16 on ROCm GPUs without native bf16 (RDNA2 and older, Vega) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12676
* fix(studio): discover On Device GGUFs from disk before Hub by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12460
* docs(save): correct save_method values in docstrings (16bit/4bit -> merged_16bit/merged_4bit) by @simpleqt in https://github.com/unslothai/unsloth/pull/10394
* Studio: deterministic FLUX.1, Z-Image and Qwen-Image renders on sm80, sm89 and sm120 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12674
* Studio: async block prefetch for MiniMax-H3's streamed denoiser by @danielhanchen in https://github.com/unslothai/unsloth/pull/12683
* Studio: fused int8 GEMM for block-streamed image denoisers by @danielhanchen in https://github.com/unslothai/unsloth/pull/12677
* Block Swap: Stream frozen transformer blocks from host RAM so a dense model can train past the card's VRAM by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/11832
* Studio: bf16 Wan VAE decode on ROCm (gfx11 / gfx12) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12681
* Studio: split Audio into Speak, Music and Transcribe workspaces by @Etherll in https://github.com/unslothai/unsloth/pull/12600
* Studio: add the Clone page, audio inputs and saved voices by @Etherll in https://github.com/unslothai/unsloth/pull/12602
* Studio: add timestamps, speakers and exports to the Transcribe page by @Etherll in https://github.com/unslothai/unsloth/pull/12607
* Studio: add the Music studio with song, sound effect and edit modes by @Etherll in https://github.com/unslothai/unsloth/pull/12609
* MLX inference test stub: accept any arguments a zoo helper is called with by @danielhanchen in https://github.com/unslothai/unsloth/pull/12700
* Studio: add the Separate audio workflow with a synced stem mixer by @Etherll in https://github.com/unslothai/unsloth/pull/12610
* Studio: add the Edit speech page by @Etherll in https://github.com/unslothai/unsloth/pull/12606
* Studio: add a Convert audio page for voice conversion with A/B compare by @Etherll in https://github.com/unslothai/unsloth/pull/12611
* Bump Desktop crates, Docker JupyterLab, setuptools and CI pins for advisories by @danielhanchen in https://github.com/unslothai/unsloth/pull/12704
* Bump yanked chacha20 to 0.10.2 in the Desktop lockfile by @danielhanchen in https://github.com/unslothai/unsloth/pull/12712
* Rename block_swap_layers to offload_layers by @danielhanchen in https://github.com/unslothai/unsloth/pull/12705
* Studio: request stream usage from llama.cpp connections so the context bar fills by @jayzhou2309 in https://github.com/unslothai/unsloth/pull/12688
* fix: preserve shared upstream gradients in RoPE backward by @taking-lying-flat in https://github.com/unslothai/unsloth/pull/12528
* fix: preserve logits in fast cross entropy backward by @taking-lying-flat in https://github.com/unslothai/unsloth/pull/12351
* Studio: keep FlashAttention 4 inside a fullgraph compile by @danielhanchen in https://github.com/unslothai/unsloth/pull/12686
* Studio: fade Run settings at the edges it scrolls past by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12722
* Studio: panel-left / panel-right icons for open sidebars, API and home-wifi icons by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12720
* fix: allow attachment-only messages in prompt queue by @Biotrioo in https://github.com/unslothai/unsloth/pull/9215
* fix(studio): deduplicate local models by path by @Biotrioo in https://github.com/unslothai/unsloth/pull/9169
* fix(studio): serialize SQLite connection closes (#10022) by @Biotrioo in https://github.com/unslothai/unsloth/pull/10069
* fix: correct easiet typo, README grammar, and duplicate issue-template frontmatter by @simpleqt in https://github.com/unslothai/unsloth/pull/10392
* Migrate Anthropic smoke probes to v1 by @pascalandr in https://github.com/unslothai/unsloth/pull/9468
* Allow compiled decode for gpt-oss by @danielhanchen in https://github.com/unslothai/unsloth/pull/12713
* Studio: avoid duplicate llama-server option families by @Apoze in https://github.com/unslothai/unsloth/pull/9621
* Revert "Studio: add per-model custom llama.cpp INI configuration (#10783)" by @danielhanchen in https://github.com/unslothai/unsloth/pull/12725
* Studio: recount tokens when status sync adopts a resident GGUF by @claxman in https://github.com/unslothai/unsloth/pull/10467
* Chat: show the Thinking control for Ollama models that report the thinking capability by @lonexreb in https://github.com/unslothai/unsloth/pull/10043
* Studio: stop writing the date into user messages by @NilayYadav in https://github.com/unslothai/unsloth/pull/12699
* tests(version_compat): AST-based symbol checks and per-model drift coverage by @Suchitra-idu in https://github.com/unslothai/unsloth/pull/6740
* Prepare an exhausted call before the budget gate, so its replay parses by @lonexreb in https://github.com/unslothai/unsloth/pull/10274
* Studio: in-app browser panel for chat by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12347
* Support pre-registered MCP OAuth clients by @ousamabenyounes in https://github.com/unslothai/unsloth/pull/7665
* docs: add 4GB VRAM QLoRA fine-tuning guide for consumer GPUs by @KafKafrnZ in https://github.com/unslothai/unsloth/pull/7605
* feat: add support for Phi-4-multimodal-instruct by @Rishabh-git10 in https://github.com/unslothai/unsloth/pull/4278
* feat: add Gefen-X (gefenx / gefenx_muon) optimizer integration by @thad0ctor in https://github.com/unslothai/unsloth/pull/7051
* Fix attention support detection for trust_remote_code models by @Shaurya-M002 in https://github.com/unslothai/unsloth/pull/8229
* Fix connected/cloud Max Tokens fallback for MiniMax M3 by @Yudeeswaran in https://github.com/unslothai/unsloth/pull/8914
* [Feature] Reasoning effort slider / LM Studio Link Provider / Open Specific folder for project by @re4 in https://github.com/unslothai/unsloth/pull/8970
* fix: pass force_download through to snapshot_download in _get_statistics by @arthi-arumugam-git in https://github.com/unslothai/unsloth/pull/8906
* Studio: mention the leftover uv cache in uninstall by @claxman in https://github.com/unslothai/unsloth/pull/10442
* studio: install torchcodec from the PyTorch cuXXX index on Linux aarch64 by @vivekvar-dl in https://github.com/unslothai/unsloth/pull/4456
* fix(install): route Linux Intel GPU hosts away from CPU-only torch by @andomeder in https://github.com/unslothai/unsloth/pull/5274
* Studio: search, sort, multi-select and collapse for project sources by @Rafael-Silva-Oliveira in https://github.com/unslothai/unsloth/pull/8850
* feat(studio): type-to-activate search, composer and prompt inputs by @Fahad090NP in https://github.com/unslothai/unsloth/pull/10303
* Studio: load S3 audio datasets — download the audio beside its manifest and point the manifest at it by @lonexreb in https://github.com/unslothai/unsloth/pull/10050
* Studio: show the eject toast before the running-chats check by @claxman in https://github.com/unslothai/unsloth/pull/10469
* Fix expected non-streaming cancellation handling by @Apoze in https://github.com/unslothai/unsloth/pull/9616
* Studio: stop the SWA config probe from adding a phantom base model and refetching every load by @claxman in https://github.com/unslothai/unsloth/pull/10607
* Studio: delete the models Studio only discovered, support files included by @atharvgaur1845 in https://github.com/unslothai/unsloth/pull/9313
* Windows: keep uv cache under the Studio root by @Max-Reisinger in https://github.com/unslothai/unsloth/pull/9007
* Document sharing models and projects across login users by @adamsiwiec1 in https://github.com/unslothai/unsloth/pull/9363
* Wire UEmbed offset pooling and SPLADE sparse output into FastSentenceTransformer by @ysys143 in https://github.com/unslothai/unsloth/pull/9322
* Don't reuse a leftover incomplete isolated Node tree by @adamsiwiec1 in https://github.com/unslothai/unsloth/pull/9361
* Studio: bound repeated llama-server respawns after SIGKILL by @JulienJBO in https://github.com/unslothai/unsloth/pull/9689
* studio: add keyboard shortcuts for UI zoom and scaling (Cmd/Ctrl + / - / 0) by @brickheadbs-claude in https://github.com/unslothai/unsloth/pull/9647
* fix: [Bug] import unsloth failed and shows UnicodeDecodeError by @chensimian in https://github.com/unslothai/unsloth/pull/9704
* Add the UEmbed unified training loss and its fine-tuning scripts by @ysys143 in https://github.com/unslothai/unsloth/pull/9323
* Track the pinned upstream UEmbed reference behind the parity test by @ysys143 in https://github.com/unslothai/unsloth/pull/9324
* Load oci:// models via llmman serve by @ericcurtin in https://github.com/unslothai/unsloth/pull/10030
* Studio: keep Switch Back visible for a snapshot-path chat by @claxman in https://github.com/unslothai/unsloth/pull/10511
* Studio: separate active context from processed tokens by @dre4moff in https://github.com/unslothai/unsloth/pull/9528
* [CHORE] Comment out dead LoRA/GELU code paths by @elrensmin in https://github.com/unslothai/unsloth/pull/9265
* fix: [Bug] Tool_Calling does not work properly by @chensimian in https://github.com/unslothai/unsloth/pull/9706
* fix: OOM for GPT OSS 120b on 183GB of VRAM (B200) by @chensimian in https://github.com/unslothai/unsloth/pull/9703
* fix: stop false link-definition probes forcing full-document renders by @apurv-1 in https://github.com/unslothai/unsloth/pull/10688
* Studio: bake a relocatable RUNPATH into the Linux llama.cpp source build by @lonexreb in https://github.com/unslothai/unsloth/pull/12398
* Studio browser: hide tab menu items that don't apply, use the toolbar's reload icon by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12728
* Add Qwen3.5 model defaults for Studio by @ashzak in https://github.com/unslothai/unsloth/pull/5958
* Studio: pipeline linked-folder RAG ingestion by @Biotrioo in https://github.com/unslothai/unsloth/pull/9873
* Studio: find a hand-added mmproj when the GGUF repo publishes none by @yzxcj797 in https://github.com/unslothai/unsloth/pull/9365
* tests: skip two version_compat suites where the daily sweep has no torch by @Sletch in https://github.com/unslothai/unsloth/pull/12274
* feat(install): add Intel GPU (XPU/SYCL) auto-detection and llama.cpp SYCL compilation by @jischebeck in https://github.com/unslothai/unsloth/pull/9084
* Studio: show the models drive in the Live Monitor when it is not the system disk by @danielhanchen in https://github.com/unslothai/unsloth/pull/12724
* fix(studio): accessibility improvements for screen reader users (NVDA/JAWS) by @devinprater in https://github.com/unslothai/unsloth/pull/6601
* Studio: open a pasted-text upload without its <pasted_text> wrapper by @MohammadHijjawi97 in https://github.com/unslothai/unsloth/pull/12198
* Studio: bound the code highlighter cache for tool cells, previews and READMEs by @claxman in https://github.com/unslothai/unsloth/pull/10609
* Studio: resolve GET /v1/models/<id>:<quant> for an on-disk quant by @danielhanchen in https://github.com/unslothai/unsloth/pull/12727
* studio: report why launcher recovery failed, not that it was absent by @Sletch in https://github.com/unslothai/unsloth/pull/9924
* CLI: show a thinking model's reasoning with --think when attached to a running Unsloth by @breken-ai in https://github.com/unslothai/unsloth/pull/11956
* Studio: show a JSON seed's values as written in the Data Recipe preview by @MohammadHijjawi97 in https://github.com/unslothai/unsloth/pull/12180
* Studio: return a missing number in a recipe dataset page as null by @MohammadHijjawi97 in https://github.com/unslothai/unsloth/pull/12103
* Offloaded embedding: one graph under compiled inference by @danielhanchen in https://github.com/unslothai/unsloth/pull/12726
* Studio: int8 activations on T4 for Qwen-Image (2.5x per step) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12690
* studio: round the window button glyphs by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12729
* Studio: hide iq4_nl KV cache where llama.cpp runs its attention on the CPU (#6272) by @rodboev in https://github.com/unslothai/unsloth/pull/7008
* Studio: seam-free VAE tiles for Qwen-Image-2.1 on low-VRAM tiers (fixes thin horizontal / vertical lines) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12696
* studio: report a denied rename as a possible ACL fault, not a scanner by @Sletch in https://github.com/unslothai/unsloth/pull/10075
* Fix incomplete links showing [blocked] during chat streaming by @umran666 in https://github.com/unslothai/unsloth/pull/12104
* Studio: opt-in recursive scanning for custom model folders by @aditya-786 in https://github.com/unslothai/unsloth/pull/6989
* fix: do not permanently purge RAG when a project is recreated by @apurv-1 in https://github.com/unslothai/unsloth/pull/10583
* Studio: support source code files in workspace folders and project knowledge bases by @harshaygadekar in https://github.com/unslothai/unsloth/pull/10395
* Manual placement: log the --fit verdict the launch actually carries by @deepspace28 in https://github.com/unslothai/unsloth/pull/10831
* fix(studio): allow loopback CORS origins and custom origin overrides in desktop mode by @harshaygadekar in https://github.com/unslothai/unsloth/pull/10084
* Studio: accept a SKILL.md saved with a UTF-8 byte order mark by @MohammadHijjawi97 in https://github.com/unslothai/unsloth/pull/12195
* Studio: make ConvRot the Z-Image-Turbo INT8 default by @danielhanchen in https://github.com/unslothai/unsloth/pull/12612
* Studio: shared fused ConvRot act-quant kernel (L4 7.8% faster) and a G4 int8 GEMM tile table by @danielhanchen in https://github.com/unslothai/unsloth/pull/12693
* Guard OAuth providers from legacy desktop clients by @NoahJenkins in https://github.com/unslothai/unsloth/pull/8747
* Studio: fused RoPE for FLUX on ROCm (FLUX.2-klein 8% faster per image, pixel-identical) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12701
* Studio: flash attention by default for FLUX.2-klein on ROCm gfx11 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12703
* studio: keep the close glyph at its original size by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12738
* Studio: keep using an already-downloaded prequant .pt; new users get the .safetensors twin by @danielhanchen in https://github.com/unslothai/unsloth/pull/12711
* Studio: "Denoising please wait..." status with step count, and a live denoising preview at no speed cost by @danielhanchen in https://github.com/unslothai/unsloth/pull/12692
* Tell npm permission failures apart from a blocked registry (#8725) by @SergiorCode in https://github.com/unslothai/unsloth/pull/8729
* Studio: keep a list item's first paragraph on its bullet in fetched pages by @MohammadHijjawi97 in https://github.com/unslothai/unsloth/pull/12179
* studio: dump the backend's own thread stacks while a stall is in progress by @hellopahe in https://github.com/unslothai/unsloth/pull/9715
* Add unsloth install-kernels for prebuilt xformers / causal_conv1d / mamba_ssm wheels by @danielhanchen in https://github.com/unslothai/unsloth/pull/12732
* Studio: cap tool-result text at 256 KB before it reaches the model by @ItsRoy69 in https://github.com/unslothai/unsloth/pull/11430
* fix(llama-prebuilt): print the correct hint for access-denied renames (WinError 5) instead of the scanner theory by @deepspace28 in https://github.com/unslothai/unsloth/pull/10182
* feat(studio): show estimated context usage before model load by @InfoSage05 in https://github.com/unslothai/unsloth/pull/9475
* Studio: keep the exit code and the first error line when training dies by @harshaygadekar in https://github.com/unslothai/unsloth/pull/11574
* Studio: chunked VAE attention on ROCm when no fused kernel takes head dim 512 (fixes FLUX.1 2048x2048 OOM) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12702
* Studio: keep streamed int8 denoisers resident up to the measured fit by @danielhanchen in https://github.com/unslothai/unsloth/pull/12687
* Studio: tidy the message action bars, branch picker and fork icon by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12735
* unsloth: explain the real fix when nvidia-smi sees a GPU torch cannot use by @vivekvar-dl in https://github.com/unslothai/unsloth/pull/9281
* Studio: resolve trust_remote_code off the event loop in /api/inference/status by @Lwrless in https://github.com/unslothai/unsloth/pull/10830
* Studio: seam-free LTX-2.3 video decode (untiled when it fits, VRAM-sized tiles otherwise) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12698
* Windows installer: prefer uv-managed Python over global winget installs by @AryanGupta1112 in https://github.com/unslothai/unsloth/pull/7825
* Studio: add a Sandbox settings tab with a Windows MXC opt-in and one-click host preparation by @danielhanchen in https://github.com/unslothai/unsloth/pull/12215
* unsloth start vibe: launch Mistral Vibe against a running Unsloth server by @xavierpestel-ai in https://github.com/unslothai/unsloth/pull/8180
* Studio browser: annotate code on demand, no idle desktop polling, smaller page cache on low-memory machines by @danielhanchen in https://github.com/unslothai/unsloth/pull/12734
* Studio: per-block CUDA graphs under every offload mode, and compile below the offload hooks by @danielhanchen in https://github.com/unslothai/unsloth/pull/12694
* perf(studio): keep the MLX GPU warm for a bounded time after generation by @Lyxot in https://github.com/unslothai/unsloth/pull/12721
* Studio: whole-step CUDA graphs under offload on top of per-block graphs (L4 FLUX.1 16 GB 10% faster, HunyuanVideo-1.5 1 GiB lighter) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12707
* Studio: stop an IPv6 blackhole from killing the backend by @Lwrless in https://github.com/unslothai/unsloth/pull/10803
* Studio: int8 ConvRot text encoder for Qwen-Image-2.1, fp8 fallback by @danielhanchen in https://github.com/unslothai/unsloth/pull/12684
* SDPA: attend packed rows per segment instead of under a dense mask (2.9x less memory, 2.5x faster without flash-attn / xformers) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12740
* offload_layers='auto': plan for batch size 2, re-plan at Trainer init for the real batch by @danielhanchen in https://github.com/unslothai/unsloth/pull/12741
* Studio: extend automatic context compaction to MLX by @dre4moff in https://github.com/unslothai/unsloth/pull/9399
* Unsloth Desktop/Studio audio: follow-ups from testing the Clone, Transcribe, Edit, Separate and Convert pages by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12706
* Studio: experimental Int8 Prefill run setting for MLX models by @Lyxot in https://github.com/unslothai/unsloth/pull/12550
* studio: enable anthropic studio tools by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12497
* Studio: keep Wan2.2-5B's leading blocks resident under group offload by @danielhanchen in https://github.com/unslothai/unsloth/pull/12691
* fix: enable XPU support for tests with agnostic test selection by @JoshuaL3000 in https://github.com/unslothai/unsloth/pull/8319
* Run bitsandbytes NF4 Linear4bit through Unsloth's NF4 kernels by @danielhanchen in https://github.com/unslothai/unsloth/pull/12739
* Deep research contract: follow the research Delete gate into the More menu item by @danielhanchen in https://github.com/unslothai/unsloth/pull/12751
* Studio: seam-free tiled decode for every image VAE whose stock tiles fall under the floor (HunyuanImage-2.1 grid, Qwen-Image, slivers) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12736
* Live Monitor models disk: translate its labels and count every tile by @danielhanchen in https://github.com/unslothai/unsloth/pull/12749
* Recount harness: bind the loader's msgs the history restore now estimates from by @danielhanchen in https://github.com/unslothai/unsloth/pull/12747
* studiobench: delete through the reply's More menu, where #12735 moved it by @danielhanchen in https://github.com/unslothai/unsloth/pull/12748
* SDPA packed-segment test: gate the CUDA memory test on a real GPU by @danielhanchen in https://github.com/unslothai/unsloth/pull/12761
* Re-measure the Studio startup budget after the Sandbox tab and action bar rework by @danielhanchen in https://github.com/unslothai/unsloth/pull/12763
* Transcript stream: never drop a phase change behind a newer update by @danielhanchen in https://github.com/unslothai/unsloth/pull/12760
* version-compat latest lanes: install requests, stop shadowing torchcodec by @danielhanchen in https://github.com/unslothai/unsloth/pull/12769
* Provider URL validation tests: count only the lookups the code under test made by @danielhanchen in https://github.com/unslothai/unsloth/pull/12771
* Studio: use the edge Gemma 4 template for the QAT E2B/E4B repos by @PundarikakshNTripathi in https://github.com/unslothai/unsloth/pull/12715
* Studio model picker: keep rows full width with overlay scrollbars, roomier On Device columns by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12746
* MLX backend tests: answer zoo's _fusion_modules in the inference stub by @danielhanchen in https://github.com/unslothai/unsloth/pull/12783
* scan-packages baseline: re-review datasets 5.1.0's readline loop by @danielhanchen in https://github.com/unslothai/unsloth/pull/12784
* Load the compiled class when auto_model is a concrete class by @danielhanchen in https://github.com/unslothai/unsloth/pull/12757
* Studio chat: no empty bubble under annotation-only messages by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12799
* Studio: give toast notifications the composer's shadow by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12800
* install-kernels: add flash-attn and install every kernel by default by @danielhanchen in https://github.com/unslothai/unsloth/pull/12744
* Branch picker chevrons: give back the 24px hit target in the same 20px slot by @danielhanchen in https://github.com/unslothai/unsloth/pull/12781
* studio: even the gaps between the window button glyphs by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12782
* studio: keep the linux desktop window resizable by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12723
* Studio: stop the Vulkan probe from popping "Entry point not found" on old Vulkan loaders by @oobabooga in https://github.com/unslothai/unsloth/pull/12762
* Frontend tests: read the composer shadow and palette seeds where #12800 left them by @danielhanchen in https://github.com/unslothai/unsloth/pull/12810
* Studio: send blank seed cells to Data Recipe prompts as empty text by @NilayYadav in https://github.com/unslothai/unsloth/pull/12789
* Studio: make adding files to a knowledge base discoverable from the chat by @oobabooga in https://github.com/unslothai/unsloth/pull/12779
* Studio: keep line breaks and ticked boxes in attached Word files by @NilayYadav in https://github.com/unslothai/unsloth/pull/12788
* chore: pre-commit autoupdate by @pre-commit-ci[bot] in https://github.com/unslothai/unsloth/pull/12780
* Studio: open Audio from More and keep it in the sidebar while it is open by @Etherll in https://github.com/unslothai/unsloth/pull/12807
* Studio: show the close button in the mobile sidebar by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12806
* Studio: 12px code font, context usage ring and small sizing fixes by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12805
* Studio: keep project folders on the Docker volume by @NilayYadav in https://github.com/unslothai/unsloth/pull/12792
* Studio: send an explicit CFG for every FLUX family on the sd.cpp engine by @danielhanchen in https://github.com/unslothai/unsloth/pull/12755
* Use a native template icon in the macOS menu bar by @guus6457 in https://github.com/unslothai/unsloth/pull/10196
* Studio: keep a trailing slash on Windows CUDA toolkit roots by @173787247 in https://github.com/unslothai/unsloth/pull/9966
* Studio: chat with and export Mac LoRAs trained on base models by @NilayYadav in https://github.com/unslothai/unsloth/pull/12794
* Studio: CI guard that fails when any family's VAE decodes in tiles below the seam floor by @danielhanchen in https://github.com/unslothai/unsloth/pull/12766
* Studio: give API clients the same Qwen thinking sampling as Chat by @NilayYadav in https://github.com/unslothai/unsloth/pull/12791
* Studio: route Qwen-Image-Edit / Edit-2509 / Edit-2511 and FLUX.1-Kontext GGUFs to the native sd.cpp engine by @danielhanchen in https://github.com/unslothai/unsloth/pull/12776
* Studio: Qwen-Image-Layered on the diffusers engine, GGUF included by @danielhanchen in https://github.com/unslothai/unsloth/pull/12775
* Studio: train Continued Pretraining on the body text column by @NilayYadav in https://github.com/unslothai/unsloth/pull/12790
* Studio: Qwen-Image-Layered on the native sd.cpp engine by @danielhanchen in https://github.com/unslothai/unsloth/pull/12777
* Studio: save audio clips and stems through the desktop save dialog by @Etherll in https://github.com/unslothai/unsloth/pull/12808
* Thread browser harnesses: reach Delete through the More menu, and attack Copy in the dismissal probe by @danielhanchen in https://github.com/unslothai/unsloth/pull/12827
* Studio: size seam-free VAE tiles from measured peaks (no slower than stock, no low-VRAM OOM) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12764
* Studio: leave MiniMax-H3's unpinned streamed blocks to diffusers' onload by @oobabooga in https://github.com/unslothai/unsloth/pull/12753
* Studio: add Manage files to Settings > Data and reorder Settings tabs by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12825
* Studio: dark dropdown glow, menu ticks and picker polish by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12826
* Run narrow RMSNorm rows several per program, bit-identical (Qwen3 q/k norms) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12754
* Studio: dismiss the llama.cpp update toast once the job finishes by @indrajeetapache in https://github.com/unslothai/unsloth/pull/9199
* fast_inference for Qwen3.5 / 3.6 MoE and Gemma-4 MoE with LoRA on the experts by @danielhanchen in https://github.com/unslothai/unsloth/pull/12742
* Studio: let safetensors vision models see every image in a chat by @NilayYadav in https://github.com/unslothai/unsloth/pull/12787
* Improve Korean Studio translations by @Jiye-Park01 in https://github.com/unslothai/unsloth/pull/9927
* Studio: add custom validator blocks to Data Recipes by @kentwait in https://github.com/unslothai/unsloth/pull/7784
* feat(studio): add model picker and reload shortcut to API monitor by @InfoSage05 in https://github.com/unslothai/unsloth/pull/11223
* Fix Windows MXC runtime grants for uv-managed Python by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12756
* Fix MXC inherited runtime permissions and pending grant recovery by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12758
* Studio: keep a Library file's star, folder and name after an edit by @NilayYadav in https://github.com/unslothai/unsloth/pull/12793
* feat(studio): opt-in reasoning for Custom Chat Completions connections by @wasimysaid in https://github.com/unslothai/unsloth/pull/12743
* Studio: make the OS sandbox work in Colab and other containers by @danielhanchen in https://github.com/unslothai/unsloth/pull/12801
* Studio: build each Qwen-Image-Edit variant on its own transformer config (2509 / original Edit no longer get 2511's zero_cond_t) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12815
* Studio: load ComfyUI-format int8_convrot / fp8 single-file DiTs instead of rendering noise by @danielhanchen in https://github.com/unslothai/unsloth/pull/12765
* Clear the duplicate-definition backlog and gate it at zero by @rajarshidattapy in https://github.com/unslothai/unsloth/pull/9769
* Studio: open chat HTML in the Browser and remove Canvas by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12804
* Keep EmbeddingGemma and Qwen3-Embedding prompts when fine-tuning by @NilayYadav in https://github.com/unslothai/unsloth/pull/12795
* fix(studio): resolve UUID-form CUDA_VISIBLE_DEVICES to physical GPU indices by @rsd-darshan in https://github.com/unslothai/unsloth/pull/8917
* Wait on the pooled fetch itself in the closed-request browser test by @danielhanchen in https://github.com/unslothai/unsloth/pull/12835
* Studio: run Qwen-Image-2.1 edits on 16 GB cards instead of refusing them by @oobabooga in https://github.com/unslothai/unsloth/pull/12752
* Fix managed vLLM startup in Windows Studio by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12750
* Studio: add Swedish locale by @yeager in https://github.com/unslothai/unsloth/pull/10071
* Studio: share one kernel install core with unsloth install-kernels by @danielhanchen in https://github.com/unslothai/unsloth/pull/12828
* Seq2Seq LoRA task type and GA token count for T5Gemma2 by @daruoktab in https://github.com/unslothai/unsloth/pull/7795
* Studio: compact long chats on API models too by @NilayYadav in https://github.com/unslothai/unsloth/pull/12786
* README: document Homebrew installation on macOS by @SSakutaro in https://github.com/unslothai/unsloth/pull/9834
* Studio: Sandbox Low/High and Permissions under Settings > Sandbox by @danielhanchen in https://github.com/unslothai/unsloth/pull/12802
* Studio: load the ComfyUI-format twin of a hosted image prequant, and run ComfyUI fp8 on the fp8 path by @danielhanchen in https://github.com/unslothai/unsloth/pull/12820
* Studio: do not offer a local model that is short a shard by @sts-change in https://github.com/unslothai/unsloth/pull/8985
* Studio: fix audio mic choice, reload page, tour copy and palette search by @Etherll in https://github.com/unslothai/unsloth/pull/12813
* Studio: discover models installed through oMLX by @sts-change in https://github.com/unslothai/unsloth/pull/8937
* Send maskless causal flex calls to SDPA is_causal by @danielhanchen in https://github.com/unslothai/unsloth/pull/12773
* Studio: more browser settings by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12829
* Studio: list every runnable audio.cpp-gguf model on its Audio pages by @Etherll in https://github.com/unslothai/unsloth/pull/12818
* Studio: add Hebrew (he) display language by @start-life in https://github.com/unslothai/unsloth/pull/6344
* Pin qwen-image-layered's VAE config for the seam guard, and check coverage in Backend CI by @danielhanchen in https://github.com/unslothai/unsloth/pull/12843
* Studio: send audio clips, stems and transcripts to other Audio pages by @Etherll in https://github.com/unslothai/unsloth/pull/12811
* Studio: import Cursor, Claude Code and Codex conversations from Settings by @Tamsi in https://github.com/unslothai/unsloth/pull/8561
* Studio: close the MLX context compaction gaps left after #9399 by @Lyxot in https://github.com/unslothai/unsloth/pull/12823
* Studio browser: mark the panel's fetches to unsloth.ai for a Cloudflare skip rule by @danielhanchen in https://github.com/unslothai/unsloth/pull/12834
* Studio: say when the audio runtime is not the release Studio installs by @Etherll in https://github.com/unslothai/unsloth/pull/12809
* Call create_optimizer without a model in the embedding optim-bits test by @vineethsaivs in https://github.com/unslothai/unsloth/pull/12697
* Runtime encoding lint: exempt the reviewed PyAV codec open by @danielhanchen in https://github.com/unslothai/unsloth/pull/12831
* Studio: tighten the browser address bar and clear it while typing by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12852
* Studio: clone with 13 more audio.cpp models, convert with Tone-Color VC by @Etherll in https://github.com/unslothai/unsloth/pull/12822
* Studio: sample Qwen-Image-2.1 and FLUX.1 dev at ComfyUI's fixed sigma schedule by @danielhanchen in https://github.com/unslothai/unsloth/pull/12838
* Studio: fix locale parity on main (Swedish sandbox strings, Manage files) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12844
* Build flex attention masks without graph breaks under torch.compile by @danielhanchen in https://github.com/unslothai/unsloth/pull/12837
* Studio: open audio models on their Audio page by @Etherll in https://github.com/unslothai/unsloth/pull/12816
* Studio: fine-tune Laya and Cloudflare Clef decision models and serve them by @NilayYadav in https://github.com/unslothai/unsloth/pull/12585
* Studio: sandbox level picker and Settings > Sandbox cleanup by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12855
* Studio: show audio clip details, playback and runs in the Library by @Etherll in https://github.com/unslothai/unsloth/pull/12814
* install.ps1: guard the mirror env restore and run its test in Windows CI by @danielhanchen in https://github.com/unslothai/unsloth/pull/12849
* Studio: size host RAM from the container's cgroup, and re-measure the MiniMax-H3 floor by @danielhanchen in https://github.com/unslothai/unsloth/pull/12854
* tests: make the torchcodec provenance and mirror repair tests hermetic by @danielhanchen in https://github.com/unslothai/unsloth/pull/12850
* Studio Desktop: stop the launch-time PATH probe from rewriting shell history by @danielhanchen in https://github.com/unslothai/unsloth/pull/12846
* Studio: complete the OpenAI-compatible audio API by @Etherll in https://github.com/unslothai/unsloth/pull/12812
* Read the API monitor unload button's disabled terms, not its exact spelling by @danielhanchen in https://github.com/unslothai/unsloth/pull/12858
* Studio: add audio translations and list audio workflows in /v1/models by @Etherll in https://github.com/unslothai/unsloth/pull/12817
* Studio: read chat replies aloud in a saved Audio voice by @Etherll in https://github.com/unslothai/unsloth/pull/12819
* Studio: convert the decoded image to uint8 on the device by @danielhanchen in https://github.com/unslothai/unsloth/pull/12847
* Studio: keep GLM-5.3-Flash's MTP head in Auto by @danielhanchen in https://github.com/unslothai/unsloth/pull/12833
* Do not print the 16bit LoRA notice when a quantization_config still quantizes by @danielhanchen in https://github.com/unslothai/unsloth/pull/12848
* Compile the Laya encoder for decision model training by @danielhanchen in https://github.com/unslothai/unsloth/pull/12778
* EmbeddingGemma 2 support and embedding improvements by @danielhanchen in https://github.com/unslothai/unsloth/pull/12865
* Studio: stop loading EmbeddingGemma in float16 in the RAG embedder by @oobabooga in https://github.com/unslothai/unsloth/pull/12864
* Studio UI test: drive the sandbox level picker that replaced the menu switch by @danielhanchen in https://github.com/unslothai/unsloth/pull/12866

## New Contributors
* @kei-yamazaki made their first contribution in https://github.com/unslothai/unsloth/pull/12475
* @1fanwang made their first contribution in https://github.com/unslothai/unsloth/pull/12505
* @simpleqt made their first contribution in https://github.com/unslothai/unsloth/pull/10394
* @jayzhou2309 made their first contribution in https://github.com/unslothai/unsloth/pull/12688
* @KafKafrnZ made their first contribution in https://github.com/unslothai/unsloth/pull/7605
* @Rishabh-git10 made their first contribution in https://github.com/unslothai/unsloth/pull/4278
* @thad0ctor made their first contribution in https://github.com/unslothai/unsloth/pull/7051
* @Shaurya-M002 made their first contribution in https://github.com/unslothai/unsloth/pull/8229
* @Yudeeswaran made their first contribution in https://github.com/unslothai/unsloth/pull/8914
* @re4 made their first contribution in https://github.com/unslothai/unsloth/pull/8970
* @arthi-arumugam-git made their first contribution in https://github.com/unslothai/unsloth/pull/8906
* @vivekvar-dl made their first contribution in https://github.com/unslothai/unsloth/pull/4456
* @andomeder made their first contribution in https://github.com/unslothai/unsloth/pull/5274
* @Rafael-Silva-Oliveira made their first contribution in https://github.com/unslothai/unsloth/pull/8850
* @Fahad090NP made their first contribution in https://github.com/unslothai/unsloth/pull/10303
* @atharvgaur1845 made their first contribution in https://github.com/unslothai/unsloth/pull/9313
* @Max-Reisinger made their first contribution in https://github.com/unslothai/unsloth/pull/9007
* @adamsiwiec1 made their first contribution in https://github.com/unslothai/unsloth/pull/9363
* @ysys143 made their first contribution in https://github.com/unslothai/unsloth/pull/9322
* @JulienJBO made their first contribution in https://github.com/unslothai/unsloth/pull/9689
* @brickheadbs-claude made their first contribution in https://github.com/unslothai/unsloth/pull/9647
* @chensimian made their first contribution in https://github.com/unslothai/unsloth/pull/9704
* @ericcurtin made their first contribution in https://github.com/unslothai/unsloth/pull/10030
* @dre4moff made their first contribution in https://github.com/unslothai/unsloth/pull/9528
* @elrensmin made their first contribution in https://github.com/unslothai/unsloth/pull/9265
* @jischebeck made their first contribution in https://github.com/unslothai/unsloth/pull/9084
* @devinprater made their first contribution in https://github.com/unslothai/unsloth/pull/6601
* @aditya-786 made their first contribution in https://github.com/unslothai/unsloth/pull/6989
* @NoahJenkins made their first contribution in https://github.com/unslothai/unsloth/pull/8747
* @SergiorCode made their first contribution in https://github.com/unslothai/unsloth/pull/8729
* @AryanGupta1112 made their first contribution in https://github.com/unslothai/unsloth/pull/7825
* @xavierpestel-ai made their first contribution in https://github.com/unslothai/unsloth/pull/8180
* @PundarikakshNTripathi made their first contribution in https://github.com/unslothai/unsloth/pull/12715
* @guus6457 made their first contribution in https://github.com/unslothai/unsloth/pull/10196
* @173787247 made their first contribution in https://github.com/unslothai/unsloth/pull/9966
* @Jiye-Park01 made their first contribution in https://github.com/unslothai/unsloth/pull/9927
* @kentwait made their first contribution in https://github.com/unslothai/unsloth/pull/7784
* @yeager made their first contribution in https://github.com/unslothai/unsloth/pull/10071
* @daruoktab made their first contribution in https://github.com/unslothai/unsloth/pull/7795
* @SSakutaro made their first contribution in https://github.com/unslothai/unsloth/pull/9834
* @start-life made their first contribution in https://github.com/unslothai/unsloth/pull/6344
* @Tamsi made their first contribution in https://github.com/unslothai/unsloth/pull/8561

**Full Changelog**: https://github.com/unslothai/unsloth/compare/v0.1.902-beta...v0.1.903-beta