v0.1.902-beta

This release brings faster navigation, shareable run settings, and clearer errors to Unsloth Desktop. It also 4x speeds up Laya decisions, expands hosted Decision API support, and keeps NVFP4, INT4, and MXFP4 checkpoints in 4-bit during LoRA training.

## Highlights

- **Command palette** on Cmd/Ctrl+P. Also, you can now share your GGUF model's run settings
- **Laya decisions up to 4.1x faster**, plus connections to TypeSafe and other hosted providers
- **Clearer errors** that explain failed loads and generations, with a View logs button
- **NVFP4 LoRA with 45% lower peak memory** on Qwen3.8-27B-NVFP4
- **INT4 + MXFP4 checkpoints stay packed** during LoRA, including Kimi-K2.7-Code and gpt-oss

## Desktop + chat

- **Cmd/Ctrl+P** opens a command palette, and GGUF run settings can be shared as a link or saved without loading.
- **Continue response** extends a finished reply, and **Fix with the model** puts HTML canvas errors in the message box.
- A warning flags models that only partly fit on the GPU, and the Images gallery loads thumbnails first.
- The desktop app adds Cmd/Ctrl + and - zoom, Windows 11 style buttons on Windows and Linux, and a check before Update stops training.

## Lower-memory checkpoint training

Pre-quantized 4-bit checkpoints now train on their exact published weights. Training reads them packed, so a loaded model needs about as much memory as its 4-bit checkpoint.

- [Qwen3.8-27B-NVFP4](https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4) peaks at 40.2 GB instead of 72.9 GB on one RTX PRO 6000, with 11% shorter steps.
- Dense compressed-tensors INT4 checkpoints train as published on NVIDIA GPUs. Qwen3-8B w4a16 peaks at 8.5 GB, not 18.8 GB, on a B200.
- Kimi-K2.7-Code trains on 6 B200s with its INT4 experts packed, where a 16-bit copy would need about 2 TB.
- gpt-oss loaded with `load_in_4bit = False` keeps its MXFP4 experts, so gpt-oss-120b trains on one B200.

**How it works**

- Triton kernels dequantize each layer on the fly, bit for bit matching compressed-tensors, and backward keeps only the packed weights.
- INT4 stays exact instead of being re-quantized to NF4, which rounds every weight twice and raised Qwen3-8B perplexity by 1.6%.
- For inference-only gpt-oss, `UNSLOTH_MXFP4_KEEP_PACKED=0` restores the native load, which prefills faster.

## Decision API + Laya

[Laya decision models](https://unsloth.ai/docs/models/decision-laya) now answer short requests up to 4.1x faster, and hosted decision models work too.

- Add TypeSafe or another hosted decision provider under Connections, then pick its models in **Settings > API**.
- With the Decision API on, enable **Unsloth Decisions** in the chat's MCP menu for a decide tool.
- On a B200, a 3-question guardrail check takes 2.8 ms instead of 11.6 ms. Long inputs run at about the same speed.
- Laya loads in about 9 s instead of 17 s and needs about half the peak host RAM.

## Image + video

- Image models keep INT8 or FP8 when offloading. On a 10 GiB-capped B200, Z-Image takes 2.3 s, not 38.4 s.
- Qwen-Image-2.1 renders 1536x1536 images up to 3.6x faster on a Radeon 8060S.

## Training + models

- Fine-tune T5, T5Gemma, BART, Marian and other text encoder-decoders with `FastModel`.
- The new `unsloth eval` command runs lm-evaluation-harness tasks on a checkpoint or LoRA adapter.
- Gemma 4 31B training and batched generation now work on models split across GPUs.

## AMD + Apple Silicon

- On Windows, RX 5000 through RX 9000 cards get PyTorch from AMD's multi-arch index.
- MLX moves to mlx 0.32.3 and mlx-vlm 0.7.4, and quantized-KV vision loads can batch replies.
- On a Mac, export LoRA adapters as PEFT or GGUF for GPU machines.

## What's Changed
* Bump install.sh / install.ps1 pin to unsloth>=2026.9.12, unsloth-zoo>=2026.9.8 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12209
* Give the Kimi processor test's stand-in tokenizer convert_tokens_to_ids by @danielhanchen in https://github.com/unslothai/unsloth/pull/12206
* Follow chat export calls in the cancelled-save contract instead of a fixed file list by @danielhanchen in https://github.com/unslothai/unsloth/pull/12203
* Add Idefics3 support (Granite Docling VLM) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4241
* Repair the checks red on main after today's Mistral-format, GKD, laya, llm-compressor and HF token merges by @danielhanchen in https://github.com/unslothai/unsloth/pull/12218
* Load compressed-tensors packed INT4 checkpoints for QLoRA without a 16-bit copy (exact packed INT4 by default, NF4 fallback), and train remote DeepSeek-V3 MoE code (Kimi-K2.7-Code) by @danielhanchen in https://github.com/unslothai/unsloth/pull/11537
* fix(studio): prevent terminal tool hangs in credential scan (#12048) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/12087
* Default loader.py's Mistral-format names in the precision-conflict harness by @danielhanchen in https://github.com/unslothai/unsloth/pull/12229
* Studio: say why a generation or a model load failed, and offer the log by @danielhanchen in https://github.com/unslothai/unsloth/pull/8804
* Train Kimi-K3 with its MXFP4 experts kept packed and dequantized on the fly by @danielhanchen in https://github.com/unslothai/unsloth/pull/11750
* fix: align GKD Liger loss with the native objective by @taking-lying-flat in https://github.com/unslothai/unsloth/pull/12007
* Studio: resolve repo-relative image paths in the simple image+text VLM converter by @breken-ai in https://github.com/unslothai/unsloth/pull/11897
* fix(studio): reuse hardware monitor drag and resize handling for API monitor by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12204
* Unsloth Studio (AMD/Windows): real torchao INT8 / FP8 export on Windows ROCm by @danielhanchen in https://github.com/unslothai/unsloth/pull/7102
* Studio: clamp unsupported reasoning effort to offered levels by @oobabooga in https://github.com/unslothai/unsloth/pull/12208
* Studio: load Laya without randomly initialising its vocabulary embedding by @danielhanchen in https://github.com/unslothai/unsloth/pull/12210
* Scope the runtime encoding scan past vendored laya while its loader supplies UTF-8 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12230
* Studio: load Images strip tiles as thumbnails instead of full PNGs by @oobabooga in https://github.com/unslothai/unsloth/pull/12228
* Images: smaller prompt box radius, inset scrollbar, taller default by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12223
* Library delete dialog: no dark mode border, cancel on outside click by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12222
* Studio: laya download confirm by @NilayYadav in https://github.com/unslothai/unsloth/pull/12196
* feat(studio): default the model source to ModelScope where Hugging Face is restricted by @Lyxot in https://github.com/unslothai/unsloth/pull/12197
* Studio: fix prompt timestamp layout, accessibility, and imported times by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12173
* Studio: explain failed and recurring llama.cpp runtime repairs by @umran666 in https://github.com/unslothai/unsloth/pull/12106
* Studio: faster Laya inference (marker-only head, CUDA graphs) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12224
* Library: equal size chat cards and no loading placeholders by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12221
* Run the user message timestamp Playwright check in Frontend CI by @danielhanchen in https://github.com/unslothai/unsloth/pull/12255
* Studio: say when a model only partly fits on the GPU by @danielhanchen in https://github.com/unslothai/unsloth/pull/8812
* Studio: push only the exported merged model to the Hub on a Mac by @NilayYadav in https://github.com/unslothai/unsloth/pull/12239
* Fix sidebar More menu hover and collapsed border hover by @Sneakr in https://github.com/unslothai/unsloth/pull/12216
* Studio: keep earlier messages when you send audio by @NilayYadav in https://github.com/unslothai/unsloth/pull/12244
* Studio: show the model a phone photo the right way up by @NilayYadav in https://github.com/unslothai/unsloth/pull/12241
* Studio: name the llama.cpp graph-scheduler abort and stop reloading into it by @danielhanchen in https://github.com/unslothai/unsloth/pull/6741
* Fix NaN gradients from compiled flex attention on short static training batches by @danielhanchen in https://github.com/unslothai/unsloth/pull/12207
* Studio: stop chats with images from blocking other chats on GGUF models by @NilayYadav in https://github.com/unslothai/unsloth/pull/12236
* Studio: stop tool calls from getting stuck in Running forever by @NilayYadav in https://github.com/unslothai/unsloth/pull/12234
* Studio: read a remote image URL on safetensors and MLX models instead of ignoring it by @NilayYadav in https://github.com/unslothai/unsloth/pull/12243
* Composer attachments: first page previews, Library icons, lighter cards by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12261
* Keep import unsloth working when vLLM needs transformers 5 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12268
* Fall back to PEFT's LoRA forward for float32 base weights by @danielhanchen in https://github.com/unslothai/unsloth/pull/7020
* Reuse the dY @ B.t() product in the fused LoRA backward by @danielhanchen in https://github.com/unslothai/unsloth/pull/12212
* Studio: let the chat model use the Decision API through MCP by @NilayYadav in https://github.com/unslothai/unsloth/pull/12232
* Read the llama.cpp changelog after it settles, by content rather than innerText by @danielhanchen in https://github.com/unslothai/unsloth/pull/12273
* Average GRPO gradients across DDP ranks, and keep MiniCPM3's pre-head scaling by @danielhanchen in https://github.com/unslothai/unsloth/pull/12193
* Add Qwen2.5 Coder model support to registry and chat templates by @danielhanchen in https://github.com/unslothai/unsloth/pull/4221
* Studio: do not fail a cached dataset load on a bookkeeping write by @BardiaKoopah in https://github.com/unslothai/unsloth/pull/8057
* Wait for the model-config Load to finish before the test moves on by @danielhanchen in https://github.com/unslothai/unsloth/pull/12270
* Scale the attachment preview text with the UI font size by @danielhanchen in https://github.com/unslothai/unsloth/pull/12276
* Unsloth Studio: explain why Model Memory skips the RAM lock when a model is fully on the GPU by @danielhanchen in https://github.com/unslothai/unsloth/pull/9571
* Studio: title a chat with the connection that answered it, not the local model by @Lyxot in https://github.com/unslothai/unsloth/pull/9106
* Fix packed and grouped Q2 GGUF variant detection by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7033
* Improve empty text dataset validation by @Imagineer99 in https://github.com/unslothai/unsloth/pull/6646
* Studio: fit the context again when MTP is forced on a reload by @NilayYadav in https://github.com/unslothai/unsloth/pull/12231
* Studio: show an error when Gemini fails a reply partway through by @NilayYadav in https://github.com/unslothai/unsloth/pull/12242
* ci: enforce npm ci on Studio install paths (follow-up to #5604) by @danielhanchen in https://github.com/unslothai/unsloth/pull/5616
* Studio: let the dataset check accept chat columns that training already supports by @NilayYadav in https://github.com/unslothai/unsloth/pull/11849
* Unsloth Studio / Desktop: allow text-encoder nvfp4 on pre-Blackwell GPUs and explain TE precision refusals by @mahiatlinux in https://github.com/unslothai/unsloth/pull/11413
* refactor(trainer): add warning for ignored eval_steps by @danielhanchen in https://github.com/unslothai/unsloth/pull/4225
* fix: handle zero-strided tensors in fast_rope_embedding (#3781) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4233
* Studio: hide the stray scrollbar under the composer at fractional zoom by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12279
* Unsloth Studio / Desktop (AMD): "Automatic" llama.cpp follows the installed torch on a mixed NVIDIA+AMD host by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12250
* Unsloth Studio / Desktop (AMD): explain why an NVIDIA+AMD host installed with `CUDA_VISIBLE_DEVICES=""` shows no GPU by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12249
* Studio desktop: let the sidebar shrink to 224px by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12283
* Studio: reopen the artifact panel after it has been dragged shut by @danielhanchen in https://github.com/unslothai/unsloth/pull/12278
* Studio: say when an attached PDF has no readable text by @NilayYadav in https://github.com/unslothai/unsloth/pull/12237
* Studio: read attached HTML files in the encoding they declare by @NilayYadav in https://github.com/unslothai/unsloth/pull/12235
* Decline the fused LoRA kernels under FSDP, where the weights are shard views by @danielhanchen in https://github.com/unslothai/unsloth/pull/11132
* Studio: allow long inputs for embedding models by @NilayYadav in https://github.com/unslothai/unsloth/pull/11721
* feat(mlx): support DDP in CLI training by @Lyxot in https://github.com/unslothai/unsloth/pull/7166
* Studio: save the chat template a base model was trained with by @NilayYadav in https://github.com/unslothai/unsloth/pull/12240
* Unsloth Studio (AMD RDNA 1, Windows): install ROCm torch for the RX 5700 XT from AMD's multi-arch index by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/11755
* Studio: keep pop-up messages off the Run settings panel by @NilayYadav in https://github.com/unslothai/unsloth/pull/11486
* Fix DDP "marked ready twice" for VLMs with CPU offload + TiledMLP by @danielhanchen in https://github.com/unslothai/unsloth/pull/4240
* Studio: show download progress during desktop updates by @NilayYadav in https://github.com/unslothai/unsloth/pull/11291
* Unsloth Studio / Desktop: add Canvas's console errors, and offer them to the model by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/11744
* fix(prompt storage): follow up, restore prompt list bookmarking, the kind pill and drag autoscroll by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/9788
* Raise the minimum typer version so the unsloth command starts by @NilayYadav in https://github.com/unslothai/unsloth/pull/11148
* Unsloth Studio: run fp16 Laya checkpoints in fp16 on MLX by @Lyxot in https://github.com/unslothai/unsloth/pull/12256
* Qwen3.5: skip compiled regions on eager decode steps, make CUDA graph decode opt-in by @danielhanchen in https://github.com/unslothai/unsloth/pull/12213
* Studio: stop the stall watchdog from killing Xet downloads that are still receiving data by @NilayYadav in https://github.com/unslothai/unsloth/pull/11594
* Studio: keep what you typed with a document when a long chat is compacted by @NilayYadav in https://github.com/unslothai/unsloth/pull/12238
* Train LoRA on NVFP4 compressed-tensors checkpoints with the weights kept packed (W4A16) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12185
* Studio: keep working when GitHub rate limits or is down by @NilayYadav in https://github.com/unslothai/unsloth/pull/10670
* Studio tests: forget the managed provider URL setting cache around each test by @danielhanchen in https://github.com/unslothai/unsloth/pull/12288
* Unsloth Studio / Desktop: show the GPU llama.cpp runs on when it is CUDA or ROCm and differs from training by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12251
* Unsloth Studio / Desktop: support custom vision projectors by @mahiatlinux in https://github.com/unslothai/unsloth/pull/11352
* Studio: Respect the thinking budget on /v1/messages by @NilayYadav in https://github.com/unslothai/unsloth/pull/11724
* Keep a model's no-placement parameters on CPU (Qwen3.8-Flash-Next n-gram table) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12141
* Studio: keep unsloth start alive while a LoRA adapter's base model downloads by @NilayYadav in https://github.com/unslothai/unsloth/pull/10550
* Baseline the thirteen unsloth-zoo 2026.9.8 findings after review by @danielhanchen in https://github.com/unslothai/unsloth/pull/12292
* Studio: show attachments when editing a message by @NilayYadav in https://github.com/unslothai/unsloth/pull/11054
* Unsloth Studio installer (AMD/Windows): install PyTorch for RX 5000 to 9000 cards from AMD's multi-arch index by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/11846
* Fix: Support past_key_values in model.generate for multi-turn conversations by @danielhanchen in https://github.com/unslothai/unsloth/pull/4232
* Studio: command palette (Cmd/Ctrl+P) by @NilayYadav in https://github.com/unslothai/unsloth/pull/6846
* Unsloth Studio / Desktop: reload llama.cpp models on reconnect by @mahiatlinux in https://github.com/unslothai/unsloth/pull/11341
* Harden timing-dependent Studio backend tests (nvidia-smi cache, cold /api/health) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12275
* Let FastModel and FastLanguageModel load the same family in one process, in either order by @danielhanchen in https://github.com/unslothai/unsloth/pull/12219
* Studio: save a model's run settings without loading it by @NilayYadav in https://github.com/unslothai/unsloth/pull/10216
* Keep LoRA on dense Linears sharing a name with fused MoE experts in PEFT's v5 conversion by @danielhanchen in https://github.com/unslothai/unsloth/pull/12281
* Train gpt-oss MXFP4 LoRA through unsloth-zoo's packed experts when load_in_16bit is not set by @danielhanchen in https://github.com/unslothai/unsloth/pull/11929
* feat(studio): reuse MLX VLM prompt snapshots in the resident vision batch by @Lyxot in https://github.com/unslothai/unsloth/pull/12263
* Let Gemma 4 31B train when it is split across GPUs by @NilayYadav in https://github.com/unslothai/unsloth/pull/12233
* Preserve repo id case when resolving a model name by @BardiaKoopah in https://github.com/unslothai/unsloth/pull/8058
* feat(studio): Mac adapter-format option for LoRA export, with GGUF adapters on macOS by @Lyxot in https://github.com/unslothai/unsloth/pull/7539
* Studio: list large MCP tool schemas compactly and load the full schema on demand by @NilayYadav in https://github.com/unslothai/unsloth/pull/11046
* Support for Seq2Seq Models (T5, T5Gemma, etc.) by @danielhanchen in https://github.com/unslothai/unsloth/pull/4226
* Studio: run large Laya requests in token-budgeted chunks by @danielhanchen in https://github.com/unslothai/unsloth/pull/12271
* Studio: stabilize generalized compare mode (clean replacement) by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7407
* Add `unsloth eval` CLI command by @NilayYadav in https://github.com/unslothai/unsloth/pull/6824
* Reload LoRA adapters trained with added tokens by @danielhanchen in https://github.com/unslothai/unsloth/pull/4219
* Unsloth Studio / Desktop: expose prefill progress in API monitor by @mahiatlinux in https://github.com/unslothai/unsloth/pull/11161
* Studio: one top row and layout fixes for the Windows desktop app by @mahiatlinux in https://github.com/unslothai/unsloth/pull/12070
* Studio desktop: Cmd/Ctrl + and - zoom with a zoom popup by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12280
* Studio: use the Hugeicons internet icon for every globe by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12321
* Studio: use FileEmpty02Icon for generic file chips by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12324
* Studio: mark adjusted backup timestamps as estimated by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12322
* Studio: stop polling focus while message menus are open by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12323
* Studio desktop tests: keep temp paths distinct when the clock repeats by @danielhanchen in https://github.com/unslothai/unsloth/pull/12296
* Studio tests: check the compare composer's paste path, not one spelling of it by @danielhanchen in https://github.com/unslothai/unsloth/pull/12325
* Tests: let the TiledMLP DDP workers import unsloth_zoo on a CPU runner by @danielhanchen in https://github.com/unslothai/unsloth/pull/12329
* Studio: use Refresh01Icon and FileEmpty02Icon everywhere by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12331
* Studio: Library grid cards keep one icon position and use the full width by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12330
* Studio: tidy the Skills dialog rows and fields by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12332
* Studio: Doc01 icon for Word and Google Docs files by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12340
* Studio: lift Library grid icons to the card's middle by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12341
* Studio: more edge padding on the model picker dropdowns by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12338
* Studio: open the sidebar More menu on hover again by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12339
* Studio: use the standard chevrons in place of Hugeicons' curved ones by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12337
* Studio: stop normal chat from opening the API monitor by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12305
* Studio: fix Java initialization and home in the Linux tool sandbox by @oobabooga in https://github.com/unslothai/unsloth/pull/12294
* Studio: don't re-prompt a finished code answer on safetensors and MLX models by @NilayYadav in https://github.com/unslothai/unsloth/pull/12309
* Studio: keep answer columns out of the system prompt by @NilayYadav in https://github.com/unslothai/unsloth/pull/12306
* unsloth start pi and dsh: stop cutting every reply at 8,192 tokens by @NilayYadav in https://github.com/unslothai/unsloth/pull/12315
* Studio: flag responses that appear to stop while quoting a token by @oobabooga in https://github.com/unslothai/unsloth/pull/12259
* Studio: keep web search source links for unsloth start claude by @NilayYadav in https://github.com/unslothai/unsloth/pull/12307
* Move the attention mask to each layer's device in the fast decode loops by @danielhanchen in https://github.com/unslothai/unsloth/pull/12290
* fix(studio): mark CSV exports as UTF-8 so Excel shows non-Latin text by @Mathews-Tom in https://github.com/unslothai/unsloth/pull/11971
* Studio: keep code indentation when an HTML file is attached to a chat by @NilayYadav in https://github.com/unslothai/unsloth/pull/12312
* Audio: keep PyAV decoding working on PyAV 19, and stop the tests needing an AMR encoder by @danielhanchen in https://github.com/unslothai/unsloth/pull/12295
* Studio: accept multiple audio files per message by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12267
* Parallel-isolation guard: exempt the nvidia-smi fake's poll deadline by @danielhanchen in https://github.com/unslothai/unsloth/pull/12353
* Studio: ask before Update stops a training run in the desktop app by @NilayYadav in https://github.com/unslothai/unsloth/pull/12313
* Validate unsloth train flags the way the config file is validated by @Abhishek-B-R in https://github.com/unslothai/unsloth/pull/11780
* Fix three main CI failures from the Mac adapter export and seq2seq PRs by @danielhanchen in https://github.com/unslothai/unsloth/pull/12354
* Studio frontend test: find the More flyout by its props in any order by @danielhanchen in https://github.com/unslothai/unsloth/pull/12356
* Studio: chunk large Qwen-Image-2.1 attention queries on ROCm by @oobabooga in https://github.com/unslothai/unsloth/pull/12335
* Studio desktop: survive AppKit exceptions during event dispatch on macOS by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12277
* Studio: fix the faint seam beside the chat composer by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12304
* feat(studio): batch KV-quantized and TurboQuant MLX loads with prompt snapshot reuse by @Lyxot in https://github.com/unslothai/unsloth/pull/12343
* Studio: load plain FP8 encoder safetensors without torchao by @oobabooga in https://github.com/unslothai/unsloth/pull/12333
* Studio: let the Mac window be moved during an app update by @NilayYadav in https://github.com/unslothai/unsloth/pull/12316
* Studio: keep menus and popovers below the desktop titlebar by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12348
* fix(studio): record the bit widths of prequantized MLX models by @Mathews-Tom in https://github.com/unslothai/unsloth/pull/11940
* Fix the ShareGPT mapping for Llama 3.1, Qwen and Gemma chat templates by @NilayYadav in https://github.com/unslothai/unsloth/pull/12314
* Studio: report CUDA graphs as on once the deferred speed profile arms them by @danielhanchen in https://github.com/unslothai/unsloth/pull/12301
* Keep packed INT4 layers off the fused inference kernel in train mode by @danielhanchen in https://github.com/unslothai/unsloth/pull/12349
* Studio: isolate the Windows Terminal on MXC's default tier by @Etherll in https://github.com/unslothai/unsloth/pull/12328
* Studio: prompt caching for OpenRouter by @Lyxot in https://github.com/unslothai/unsloth/pull/12264
* Force non-reentrant gradient checkpointing for DeepSeek-V4.1 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12318
* Studio: count tokens for templates that refuse an empty chat by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12336
* Studio: catch the tensor split abort behind a gdb backtrace by @oobabooga in https://github.com/unslothai/unsloth/pull/12293
* Keep frozen BatchNorm running stats fixed during LoRA training by @danielhanchen in https://github.com/unslothai/unsloth/pull/12319
* Widen the pinned child GPU mask when extra args name a companion device (#11810) by @deepspace28 in https://github.com/unslothai/unsloth/pull/11823
* Studio: turn train on completions back on when leaving CPT by @NilayYadav in https://github.com/unslothai/unsloth/pull/12308
* Studio: fix Qwen-Image-2.1 recompiles and Ideogram 4 attention and device errors on torch 2.11 to 2.14 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12350
* Studio: fix the desktop island corner and border seam by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12285
* Move the Unsloth Studio MLX pins to mlx 0.32.3 and mlx-vlm 0.7.4 by @Lyxot in https://github.com/unslothai/unsloth/pull/12334
* Honour FP8Linear.block_size in the patched FP8 forward (32x32 block checkpoints) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12317
* Write the Ollama Modelfile from the trained chat template by @NilayYadav in https://github.com/unslothai/unsloth/pull/12311
* Fix MiniMax-H3 compiled render on torch 2.12, 2.13 and 2.14 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12326
* Studio: load LTX-2 and LTX-2.3 on the pinned transformers 5.5 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12299
* scan_packages: refuse VCS, URL and local-path specs before pip download by @danielhanchen in https://github.com/unslothai/unsloth/pull/12357
* Studio: let GGUF vision models see dark text in transparent images by @NilayYadav in https://github.com/unslothai/unsloth/pull/12310
* Keep masked head_dim 256 SDPA training off cuDNN attention on SM100 (torch 2.14 NaN grads) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12344
* Unsloth Studio (AMD): run native image generation on an NVIDIA card next to ROCm torch by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12252
* fix(lora): preserve all active adapters with PEFT fallbacks by @taking-lying-flat in https://github.com/unslothai/unsloth/pull/12214
* Studio: load downloaded tokenizers offline, size unsloth mirrors from the family table by @danielhanchen in https://github.com/unslothai/unsloth/pull/12300
* Build the marlin_gemm call from the op schema so packed INT4 inference works on vLLM 0.29 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12320
* Studio video: keep an explicit fast LTX-2.3 single-file load resident when it fits by @danielhanchen in https://github.com/unslothai/unsloth/pull/12345
* Studio: keep MiniMax-H3's int8 denoiser and compile when it has to offload by @danielhanchen in https://github.com/unslothai/unsloth/pull/12286
* Model-config UI test: stop waiting on a request whose finish event never arrives by @danielhanchen in https://github.com/unslothai/unsloth/pull/12363
* Studio: keep INT8 and compile on conventional video models when they have to offload by @danielhanchen in https://github.com/unslothai/unsloth/pull/12289
* Studio frontend test: drain the previous runtime's writes before each reasoning-effort scenario by @danielhanchen in https://github.com/unslothai/unsloth/pull/12367
* Baseline the moved transformers 5.18.0 testing_utils polling-loop site after review by @danielhanchen in https://github.com/unslothai/unsloth/pull/12368
* Studio: keep int8 / fp8 image transformers quantised when the load has to offload by @danielhanchen in https://github.com/unslothai/unsloth/pull/12287
* Studio: style canvas notices like toasts by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12358
* Studio: keep the model name ahead of its format and quant when the header is tight by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12396
* Studio: compact context usage ring when the chat header is squeezed by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12355
* fix(device_type): flush and fence through the shared device helpers by @li-lizhe in https://github.com/unslothai/unsloth/pull/12284
* Studio: open a skill when its row chevron is clicked by @NilayYadav in https://github.com/unslothai/unsloth/pull/12375
* Studio: keep the answers typed into a PDF form when it is attached to a chat by @NilayYadav in https://github.com/unslothai/unsloth/pull/12377
* Studio: stop offering Claude sampling settings that are silently ignored by @NilayYadav in https://github.com/unslothai/unsloth/pull/12383
* Studio: count companion assets in the Hub download size of image and video GGUFs by @oobabooga in https://github.com/unslothai/unsloth/pull/12370
* Feat share model run settings through links by @Sneakr in https://github.com/unslothai/unsloth/pull/11710
* studio: make media family overrides structural by @Imagineer99 in https://github.com/unslothai/unsloth/pull/10150
* Studio: keep <placeholder> and Vec<T> text in chat replies by @NilayYadav in https://github.com/unslothai/unsloth/pull/12376
* Harden installer source selection, ROCm helper staging and npm scanner cleanup by @danielhanchen in https://github.com/unslothai/unsloth/pull/12402
* Remove xFormers built for another torch after Linux repair by @tweekli in https://github.com/unslothai/unsloth/pull/11783
* truststore: keep TLS verification on when handshakes overlap on one context by @danielhanchen in https://github.com/unslothai/unsloth/pull/12403
* Read a sign-flipped SVD basis as the same subspace in the Q-GaLore schedule by @danielhanchen in https://github.com/unslothai/unsloth/pull/12399
* Create the SentencePiece scratch directory without a check-then-create race by @danielhanchen in https://github.com/unslothai/unsloth/pull/12400
* Only turn on HF_HUB_ENABLE_HF_TRANSFER in synthetic.py when hf_transfer is installed by @danielhanchen in https://github.com/unslothai/unsloth/pull/12406
* Studio: train an uploaded CSV's NA, None and 00501 cells as written by @NilayYadav in https://github.com/unslothai/unsloth/pull/12379
* Studio: fix Base vs LoRA compare for voice messages on fine-tuned Whisper by @NilayYadav in https://github.com/unslothai/unsloth/pull/12381
* Compare Kaggle reference repo ids without case and prefetch the Qwen3 4bit repo under its new spelling by @danielhanchen in https://github.com/unslothai/unsloth/pull/12411
* Seed preview: check path-backed image cells against the account's workspace by @danielhanchen in https://github.com/unslothai/unsloth/pull/12404
* Studio: say Pin in model menus, and drop the border on right-click submenus by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12414
* Stop the Chat UI persisted-monitor reset from running script in the stale page by @danielhanchen in https://github.com/unslothai/unsloth/pull/12412
* Studio: keep a local model's built-in system prompt when the date setting is on by @NilayYadav in https://github.com/unslothai/unsloth/pull/12382
* Studio: let the Decision API use TypeSafe, Liquid AI, OpenRouter and other System One servers by @NilayYadav in https://github.com/unslothai/unsloth/pull/12373
* Keep the MLX adamw_8bit optimizer instead of collapsing it to adamw by @vineethsaivs in https://github.com/unslothai/unsloth/pull/9950
* Treat {{ and }} in a merged prompt as literal braces by @vineethsaivs in https://github.com/unslothai/unsloth/pull/10742
* fix(studio): keep scoped download progress tied to current files by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12390
* Studio: continue finished replies and resume GGUF reasoning by @oobabooga in https://github.com/unslothai/unsloth/pull/12371
* Studio: keep Qwen-Image-2.1 image conditioning finite on ROCm by @oobabooga in https://github.com/unslothai/unsloth/pull/12360
* Restore full-rank gradients after Q-GaLore updates by @vineethsaivs in https://github.com/unslothai/unsloth/pull/10878
* Studio: fix PDF previews after a PDF attachment is extracted by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12346
* Studio: stop the backend crashing on startup when memory is tight by @NilayYadav in https://github.com/unslothai/unsloth/pull/12374
* Studio: warn that a Web share link from a remote Studio exposes its address by @danielhanchen in https://github.com/unslothai/unsloth/pull/12413
* Studio: smaller scroll to bottom button, visible in dark mode, with a setting to hide it by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12417
* Stop conversation_extension crashing when the caller keeps their columns by @vineethsaivs in https://github.com/unslothai/unsloth/pull/8373
* Fix desktop icon clarity with supplied artwork and Tauri resource icons by @wasimysaid in https://github.com/unslothai/unsloth/pull/12352
* Studio: always show composer attachments as cards, rename sent layouts by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12418
* Studio: remove the Projects section setting from Chat settings by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12421
* Studio: keep the arrow cursor on a sent prompt's time by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12420
* Keep an explicit HF_HUB_ENABLE_HF_TRANSFER and install hf_transfer in Core CI by @danielhanchen in https://github.com/unslothai/unsloth/pull/12394
* Studio: remove the white seam above the chat panel on Windows dark mode by @danielhanchen in https://github.com/unslothai/unsloth/pull/12425
* Studio: avoid slow cold MIOpen searches on gfx1151 by @oobabooga in https://github.com/unslothai/unsloth/pull/12359
* Studio: faster MiniMax-H3 GGUF renders (resident under memory auto, sd.cpp pin upgrade, speed_mode=max kernels) by @danielhanchen in https://github.com/unslothai/unsloth/pull/12405
* Studio: keep Word footnotes and endnotes in chats and knowledge bases by @NilayYadav in https://github.com/unslothai/unsloth/pull/12378
* Studio: make Compare in Chat load the full fine-tune that just finished by @NilayYadav in https://github.com/unslothai/unsloth/pull/12380
* Stop the model selector's format and quant suffix clipping descenders by @danielhanchen in https://github.com/unslothai/unsloth/pull/12427
* fix: allow configuring Pi output token limit by @Imagineer99 in https://github.com/unslothai/unsloth/pull/12393
* Studio: keep the sidebar's bottom fade in step with the list by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12423
* Studio: keep the 1024 canvas when auto precision picked the Qwen-Image-2.1 quant by @danielhanchen in https://github.com/unslothai/unsloth/pull/12428
* Studio: move Blender MCP setup out of Manage MCP servers into the composer by @NilayYadav in https://github.com/unslothai/unsloth/pull/12430
* Studio: explain the all-columns-dropped recipe error in UI terms by @danielhanchen in https://github.com/unslothai/unsloth/pull/12416
* Studio: Qwen-Image-2.1 placement from measured sizes, int8 under offload on torchao 0.17, balanced fit check by @danielhanchen in https://github.com/unslothai/unsloth/pull/12408
* Read the desktop New chat button contract token by token by @danielhanchen in https://github.com/unslothai/unsloth/pull/12426
* feat(studio): configurable RAG upload extensions via RAG_UPLOAD_EXTS by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/11499
* Unsloth Studio: return the spoken text from /audio/generate instead of a truncated label by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/12386
* Studio: keep PyTorch mirror leaves out of query tokens by @alkinun in https://github.com/unslothai/unsloth/pull/10546
* perf(studio): reuse embeddings for identical files in linked folders by @Mathews-Tom in https://github.com/unslothai/unsloth/pull/11972
* Studio: render MCP Apps widgets in the chat thread by @NilayYadav in https://github.com/unslothai/unsloth/pull/9301
* Bump install.sh / install.ps1 pins to unsloth>=2026.9.13, unsloth-zoo>=2026.9.9 by @danielhanchen in https://github.com/unslothai/unsloth/pull/12432
* fix(studio): use resolved public id in embeddings/completions monitor by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/9346
* Studio: keep rounded boxes rounded when they scroll by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/12431
* Studio: browse temporary Linux mounts under /media and /mnt by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10048
* Make the registry's default quant_type usable by @vineethsaivs in https://github.com/unslothai/unsloth/pull/9887
* Size the packed attention mask at the padded length, not the token count by @vineethsaivs in https://github.com/unslothai/unsloth/pull/8278
* Keep the GRPO eval batch a multiple of num_generations by @vineethsaivs in https://github.com/unslothai/unsloth/pull/12362
* Prefetch Tauri NSIS and WebView2 tools before the Windows desktop build by @danielhanchen in https://github.com/unslothai/unsloth/pull/12434
* fix(studio): retry custom gateways with max_completion_tokens after max_tokens 400 (#10787) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/10837
* Collapse the llama.cpp install branch that never branched by @vineethsaivs in https://github.com/unslothai/unsloth/pull/10268
* Resolve a sentence-transformers modules.json module class through the same trust gate as upstream by @danielhanchen in https://github.com/unslothai/unsloth/pull/12436

## New Contributors
* @Mathews-Tom made their first contribution in https://github.com/unslothai/unsloth/pull/11971

**Full Changelog**: https://github.com/unslothai/unsloth/compare/v0.1.900-beta...v0.1.902-beta