v0.1.512-beta

Hey everyone! Kimi K3 & DeepSeek v4 Flash can now run locally with Unsloth Dynamic GGUFs, Unsloth can keep multiple chats generating in parallel, and the new Deep Research mode plans, reads and cites sources using your local model.

This release also brings better AMD and Intel GPU support, DoRA training, and many installer, MLX, export and inference fixes.

## DeepSeek V4 Flash 0731 (August 2 Update)

You can run our [DeepSeek V4 Dynamic GGUFs](https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF) through Unsloth. Unsloth automatically detects multi-GPU setups and can offload model layers to system memory.

- `UD-IQ1_S` is 83GB in disk space.
- `UD-IQ3_XXS` is 104GB in disk space and fits in 128GB machines.
- `UD-Q4_K_XL` is 155GB in disk space.
- For lossless inference, use `UD-Q8_K_XL`, which is 162TB in disk space.

Read our full [DeepSeek V4 Flash guide](https://unsloth.ai/docs/models/deepseek-v4)

## Kimi K3

Moonshot AI’s **Kimi K3** is a 2.8T-parameter MoE model with 104B active parameters, native vision support and a 1M context window. Kimi K3 is thinking-only and Unsloth supports low, high and max reasoning efforts.

You can run our [Kimi K3 Dynamic GGUFs](https://huggingface.co/unsloth/Kimi-K3-GGUF) through Unsloth. Unsloth automatically detects multi-GPU setups and offloads layers automatically.

Kimi K3 is a very large model, so plan your hardware accordingly:

- `UD-IQ1_S` is 595GB in disk space.
- `UD-Q4_K_XL` is 1.51TB in disk space.
- For lossless inference, use `UD-Q8_K_XL`, which is 1.56TB in disk space.

Read our full [Kimi K3 guide](https://unsloth.ai/docs/models/kimi-k3) and learn more about [Unsloth Dynamic 2.0 GGUFs](https://unsloth.ai/docs/basics/unsloth-dynamic-2.0-ggufs).

## XET and HTTP download improvements (August 3 Update)
Unsloth now auto fallbacks to HTTP if XET stalls and memory usage is much reduced. Makes downloading more stable and more efficient.

## Parallel Chat

Unsloth can now run multiple conversations at once. Starting a **New Chat** leaves the previous answer generating, and each active conversation gets its own progress indicator and Stop control.

- 4 llama-server slots by default, adjustable in the web UI.
- Tools, uploads, self-healing and agents stay isolated between chats.
- Stop one chat without interrupting others or restarting the server.
- Unsloth reduces the slot count when memory is limited.
- Reloading the model still stops active chats after confirmation.

## Deep Research

Deep Research turns a local model into a complete research workflow. Give it a question and it will create a plan, search the web, organize the evidence and produce a cited report.

- Review and edit the plan before research begins.
- Follow its progress and collected sources while it works.
- Resume or cancel without losing completed research.
- Allow or block websites to control source selection.
- Sources and citations stay saved with each run.
- Before writing, Unsloth checks for unsupported claims, contradictions and unresolved gaps. Recommendations without direct evidence are marked as testable inferences.

One run can be active per chat. Deep Research currently works with local models. Optional full-page grounding can read top search results into temporary RAG before synthesis; enable it with `UNSLOTH_RESEARCH_AUTO_SCRAPE=1`. It remains off by default because it adds scraping time and needs additional context.

## Improved AMD Support

This update builds on our initial [AMD release and setup guide](https://unsloth.ai/docs/basics/amd) with support for more GPUs and more reliable ROCm inference and training.

- Improved detection for RDNA2, Radeon, Ryzen, Strix Halo and workstation GPUs.
- MI50 and Radeon VII support 16-bit LoRA and full fine-tuning on Linux.
- Fixed 4-bit NaNs, library conflicts and long RDNA startup stalls.
- Unsupported Windows HIP GPUs can fall back to Vulkan.
- Vulkan devices show their real names and can be selected individually.
- Clean Windows installs no longer require Winget or developer tools.

## Training, Models and Platform Updates

- Intel XPU enables local chat and training on Intel Arc and Data Center GPUs.
- DoRA is available alongside LoRA and full fine-tuning.
- Large exports can use all visible GPUs to avoid GPU 0 memory limits.
- MLX adds streaming datasets, continued pretraining, callbacks and better VLM and LoRA support.
- Fixed incorrect `flash_attention_2` model output.
- Unsloth and saving utilities now work without bitsandbytes.
- Fixed `.json` datasets and Qwen3.5/3.6 MoE and GRPO notebook setup.
- Hub download folders are easier to find and open.
- Release notes are available in the Unsloth update popup.
- `unsloth start --as-subagent` lets Codex, Claude Code and Pi delegate tasks to a local Unsloth model.

## Updating / installing Unsloth

**macOS, Linux, WSL:**

```bash
curl -fsSL https://unsloth.ai/install.sh | sh
```

**Windows:**

```powershell
irm https://unsloth.ai/install.ps1 | iex
```

## What's Changed
* Bump install.sh / install.ps1 pin to unsloth>=2026.7.4 by @danielhanchen in https://github.com/unslothai/unsloth/pull/7263
* Studio: don't let a malformed HF token empty the model picker's Recommended list by @oobabooga in https://github.com/unslothai/unsloth/pull/7266
* Fix local CLI streamed generation error handling by @Lyxot in https://github.com/unslothai/unsloth/pull/7135
* Studio: make tab navigation feel immediate by @oobabooga in https://github.com/unslothai/unsloth/pull/7271
* Feat/model picker per model config v2 by @Sneakr in https://github.com/unslothai/unsloth/pull/7207
* Fix Studio desktop reliability by @wasimysaid in https://github.com/unslothai/unsloth/pull/7255
* Fix resume training crash recovery and MLX checkpoints by @NilayYadav in https://github.com/unslothai/unsloth/pull/6796
* Studio: fix loading split GGUFs from the local HF cache by @oobabooga in https://github.com/unslothai/unsloth/pull/7273
* Show iconless models in Hub feed above likes threshold by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7284
* Installer: report the installed Unsloth version by @oobabooga in https://github.com/unslothai/unsloth/pull/7265
* Studio: show HF token tick only after validation by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7268
* Installer: enable ROCm torch on RDNA2 (gfx1030-1036) on Windows by @danielhanchen in https://github.com/unslothai/unsloth/pull/7277
* show the mobile sidebar trigger above the chat header by @NilayYadav in https://github.com/unslothai/unsloth/pull/7267
* collapse composer pills to icons on narrow screens by @NilayYadav in https://github.com/unslothai/unsloth/pull/7269
* Add project pinning, Projects list view, and project chat-session fixes by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7291
* fix(rocm): prepend system ROCm libs on native Linux to avoid bundled HIP crash by @InfoSage05 in https://github.com/unslothai/unsloth/pull/7233
* studio: extend the _grouped_mm null-kernel guard to Linux ROCm RDNA4 (gfx1201) by @danielhanchen in https://github.com/unslothai/unsloth/pull/7292
* Recognize Radeon 8065S (Gorgon Halo / Ryzen AI Max 400) as gfx1151 by @danielhanchen in https://github.com/unslothai/unsloth/pull/7290
* fix(install.ps1): use ordinal IndexOf when stripping index URL credentials by @Solaris-star in https://github.com/unslothai/unsloth/pull/7286
* install.sh: route Strix to the AMD arch index on rocm7.2 (#7264) + PCI detection hint by @danielhanchen in https://github.com/unslothai/unsloth/pull/7293
* Studio: move sidebar search into the header by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7304
* Studio: reuse MLX prompt cache across turns instead of re-prefilling by @NilayYadav in https://github.com/unslothai/unsloth/pull/7311
* Unsloth start improvements: download progress, server reuse, and safe model switching by @oobabooga in https://github.com/unslothai/unsloth/pull/7313
* Baseline two benign unsloth-zoo test-file findings in scan_packages by @danielhanchen in https://github.com/unslothai/unsloth/pull/7325
* studio: show system-wide VRAM in the multi-GPU System tab view on ROCm by @hakanbaysal in https://github.com/unslothai/unsloth/pull/7216
* studio: classify embedding models from the HF cache and honor offline mode by @hakanbaysal in https://github.com/unslothai/unsloth/pull/7218
* Unsloth start: add local subagents for Claude Code, Codex, OpenCode and Pi by @danielhanchen in https://github.com/unslothai/unsloth/pull/7326
* fix(install): route Strix to AMD gfx index on ROCm 7.14 by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7300
* Faster safetensors weight loading on unified-memory (integrated) GPUs by @danielhanchen in https://github.com/unslothai/unsloth/pull/5988
* Studio: lighten chat text weight on Linux to match macOS rendering by @oobabooga in https://github.com/unslothai/unsloth/pull/7308
* Studio: reorder sidebar, rename Hub to Models by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7327
* Studio: mask AMD GPU pins via ROCR so an unsupported iGPU can't crash llama-server by @guerrieroriccardo in https://github.com/unslothai/unsloth/pull/7272
* fix(install): infer Strix gfx when ROCm runtime is absent by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7305
* Studio: calibrate Linux chat typography against macOS by @oobabooga in https://github.com/unslothai/unsloth/pull/7337
* Studio: prevent login error text clipping by @oobabooga in https://github.com/unslothai/unsloth/pull/7343
* Studio: fix stuck composer prompt on first send and unreachable --secure Cloudflare links by @NilayYadav in https://github.com/unslothai/unsloth/pull/7340
* installer: fix false "no GPU detected" on AMD hosts (dead KFD check) + clearer ROCm-less warning by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7314
* Studio: reject whitespace-only passwords by @NilayYadav in https://github.com/unslothai/unsloth/pull/7341
* Studio: UI font size scales all text consistently without moving layout by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7355
* Studio: guard project chat rename against IME composition keys by @danielhanchen in https://github.com/unslothai/unsloth/pull/7246
* fix(dataprep): don't emit a degenerate chunk for empty text by @chuenchen309 in https://github.com/unslothai/unsloth/pull/7183
* Studio: fix Connections settings tab overflow in the settings dialog by @danielhanchen in https://github.com/unslothai/unsloth/pull/7241
* Studio: drive UI font size through a typography scale instead of the root font size by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7359
* Installer: restore interrupted updates and clean stale rollback environments by @oobabooga in https://github.com/unslothai/unsloth/pull/7342
* Studio: add configurable model download location by @oobabooga in https://github.com/unslothai/unsloth/pull/7274
* Studio: add local speech-to-text dictation engine by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7095
* unsloth start/run: tool-call flags, positional model, and grouped help by @danielhanchen in https://github.com/unslothai/unsloth/pull/7328
* Security audit: refresh scan baselines for current dependency set by @danielhanchen in https://github.com/unslothai/unsloth/pull/7362
* Studio: always show the executed Python script in chat with a download option by @danielhanchen in https://github.com/unslothai/unsloth/pull/7240
* feat(studio): Mac-aware training controls for MLX (optimizers, LoftQ, packing) by @Lyxot in https://github.com/unslothai/unsloth/pull/7358
* Studio: viewport-gated highlight for the executed Python script (follow-up to #7240) by @danielhanchen in https://github.com/unslothai/unsloth/pull/7363
* Studio STT: only load safetensors weights for custom dictation models (RCE fix) by @danielhanchen in https://github.com/unslothai/unsloth/pull/7364
* Bump install.sh / install.ps1 pin to unsloth>=2026.7.5 by @danielhanchen in https://github.com/unslothai/unsloth/pull/7365
* Fix agent launch PWD normalization by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7110
* Studio: fix composer reset after failed send by @NilayYadav in https://github.com/unslothai/unsloth/pull/7377
* fix(studio): show chat sidebar menu on touch devices by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7297
* fix(studio): persist connection model selections for remote clients by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7298
* fix: pin torchcodec for torch 2.10 and warn on ABI mismatch by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7299
* fix(studio): opt-in source-build GPU smoke validation by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7322
* fix(studio): resolve bare git on Windows sandbox PATH by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7323
* fix(install): show detected distro in sudo apt Accept prompt by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7324
* Studio: fail closed on index-referenced nested pickle shards in the offline embedding gate by @danielhanchen in https://github.com/unslothai/unsloth/pull/7366
* Studio whisper: pair slim bundles on the ggml commit, not the full llama tag by @danielhanchen in https://github.com/unslothai/unsloth/pull/7381
* Unsloth start: keep Claude subagents on the local model by @danielhanchen in https://github.com/unslothai/unsloth/pull/7333
* Unsloth start: Fix local subagents in Codex, Claude Code and Pi by @oobabooga in https://github.com/unslothai/unsloth/pull/7329
* unsloth run/start: apply per-model recommended sampling + sampling override flags by @danielhanchen in https://github.com/unslothai/unsloth/pull/7335
* Studio: refine GGUF per-GPU selection (gpu_ids) by @danielhanchen in https://github.com/unslothai/unsloth/pull/7239
* Studio UI test: survive a renderer crash in the voice model-picker step by @danielhanchen in https://github.com/unslothai/unsloth/pull/7389
* Studio: use text-ui-* tokens instead of raw px in the voice tab by @danielhanchen in https://github.com/unslothai/unsloth/pull/7387
* enable studio for intel GPU by @leizhenyuan in https://github.com/unslothai/unsloth/pull/4724
* Studio: register text-ui tokens with tailwind-merge so cn() keeps them by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7396
* Studio: scan HF cache snapshot loads by their repo id by @danielhanchen in https://github.com/unslothai/unsloth/pull/7398
* Fix ROCm wheel-index unit test: extract the gfx-arch probe helpers get_torch_index_url now calls by @danielhanchen in https://github.com/unslothai/unsloth/pull/7399
* fix(studio): stop false MTP/vision capability reports by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7332
* feat(studio): expose full KV cache dtype list in model config UI by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7348
* fix(studio/colab): restore blank Colab iframe embed (#7344) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7349
* feat(studio): presets include load settings (#7347) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7352
* feat(studio): add DoRA support to studio by @lzvxck in https://github.com/unslothai/unsloth/pull/7315
* Studio: use Hugeicons AI Security glyph for Run automatically permission by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7409
* Fix Studio model picker toolbar overflow by @Lyxot in https://github.com/unslothai/unsloth/pull/7403
* Studio: scale menu, toast, chat and composer icons with the UI font size by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7400
* Studio: prevent empty responses after model thinking by @oobabooga in https://github.com/unslothai/unsloth/pull/7418
* Recipe Studio: full-height canvas and in-app maximize control by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7394
* Unsloth Studio (desktop): fix canvas preview, download file button, toast placement, and model-load typing lag by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7391
* Restore the inheritance-before-guard ordering assertion in the gguf order test by @danielhanchen in https://github.com/unslothai/unsloth/pull/7251
* fix(studio): honor run settings on initial model load (#7346) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7351
* Studio: prevent long reasoning from jumping the chat on completion by @oobabooga in https://github.com/unslothai/unsloth/pull/7388
* Studio: reset quantized KV cache to f16 when the flash-attn-off crash-recovery fallback fires by @danielhanchen in https://github.com/unslothai/unsloth/pull/7390
* fix(studio/colab): fix OutStream startup crash and tidy the notebook cards by @danielhanchen in https://github.com/unslothai/unsloth/pull/7404
* tests: anchor the gguf ordering assertion on the branch that owns the marker by @danielhanchen in https://github.com/unslothai/unsloth/pull/7443
* AMD: CI coverage for recent fixes, plus three wrong gfx ids by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7431
* fix(studio): support hostname-based enterprise proxies by @alkinun in https://github.com/unslothai/unsloth/pull/7416
* Settings: slightly wider and taller dialog by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7456
* Sidebar: settings gear above the profile in the collapsed rail by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7458
* Add drag and drop sources to the create project dialog by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7441
* Settings: match dialog fills to the app shell surfaces by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7457
* Add Agents settings tab for unsloth start by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7303
* Studio: fail fast on out-of-disk instead of a doomed llama.cpp source build by @NilayYadav in https://github.com/unslothai/unsloth/pull/7420
* test: fast end-to-end GRPO fast_inference vLLM rollout test by @JoshuaL3000 in https://github.com/unslothai/unsloth/pull/7136
* Studio: reset the reasoning open state when a new stream starts by @danielhanchen in https://github.com/unslothai/unsloth/pull/7444
* studio: shard export checkpoint loads across all visible GPUs by @hakanbaysal in https://github.com/unslothai/unsloth/pull/7215
* install.sh: do not assume sudo consent when there is no terminal by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7435
* Fix the CPU-only ROCm routing errors and two font-scale UI flakes by @danielhanchen in https://github.com/unslothai/unsloth/pull/7469
* Studio: recover tool-enabled GGUF chats after llama-server exits by @oobabooga in https://github.com/unslothai/unsloth/pull/7424
* Studio: use the UI font scale tokens in the Agents settings tab by @danielhanchen in https://github.com/unslothai/unsloth/pull/7462
* Studio: use the scaling text tokens in the Agents settings tab by @danielhanchen in https://github.com/unslothai/unsloth/pull/7468
* studio/tests: cover the GGUF load ordering behaviourally and make the structlog stub order-independent by @danielhanchen in https://github.com/unslothai/unsloth/pull/7442
* install.sh, setup.sh: apply the no-tty consent fix to the remaining sites by @danielhanchen in https://github.com/unslothai/unsloth/pull/7470
* Import bitsandbytes before the hardware spoof rewrites torch by @danielhanchen in https://github.com/unslothai/unsloth/pull/7471
* Fix PDF-grounded QA recipe for QLoRA by @Etherll in https://github.com/unslothai/unsloth/pull/7107
* Studio: default tool-call permission to Approve for me, prompt only on high-risk actions by @danielhanchen in https://github.com/unslothai/unsloth/pull/7285
* Add interactive Agents command builder by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7312
* Baseline the fastapi SSE keepalive loop after its 0.140.0 rewrite by @danielhanchen in https://github.com/unslothai/unsloth/pull/7480
* Bypass fast_generate for flash_attention_2 models (StaticCache + FA2 produces gibberish) by @pwasiewi in https://github.com/unslothai/unsloth/pull/7429
* fix(studio): reject Vulkan diffusion gpu_ids before Phase 1 teardown by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7415
* fix(studio): report Vulkan GPUs in system UI by @alkinun in https://github.com/unslothai/unsloth/pull/7476
* tests: read checked-in files as UTF-8 instead of the platform default by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7438
* Studio: add Deep Research by @alkinun in https://github.com/unslothai/unsloth/pull/7219
* Studio: add UNSLOTH_STUDIO_DISABLE_PUBLIC_CHECK to switch off the startup public lookups by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7433
* Fix the wall-clock timeout tests on Python 3.10 by @danielhanchen in https://github.com/unslothai/unsloth/pull/7488
* Pin utf-8 on shipping-code text I/O instead of the operator locale by @danielhanchen in https://github.com/unslothai/unsloth/pull/7486
* Let a decode failure degrade instead of escaping a fail-closed helper by @danielhanchen in https://github.com/unslothai/unsloth/pull/7487
* Remove the no-op rmtree guard around the GGUF save by @vineethsaivs in https://github.com/unslothai/unsloth/pull/7479
* Studio: fetch bare hostnames as https instead of refusing them by @NilayYadav in https://github.com/unslothai/unsloth/pull/7427
* tests: record the gfx1152 llama.cpp bundle gap so the next one is not silent by @danielhanchen in https://github.com/unslothai/unsloth/pull/7491
* unsloth start: keep the local subagent unattended and out of plan mode by @danielhanchen in https://github.com/unslothai/unsloth/pull/7437
* Keep the newer-mapper probe from replacing the installed FP8 mappers by @vineethsaivs in https://github.com/unslothai/unsloth/pull/7478
* Studio: admission control on /v1/messages, slot pool that tracks --parallel by @danielhanchen in https://github.com/unslothai/unsloth/pull/7436
* Keep the newer-mapper probe alive when the fetched mapper has no fp8 tables by @danielhanchen in https://github.com/unslothai/unsloth/pull/7497
* Studio: fix the "No model loaded" error, and optionally auto-download a model named in an API request by @danielhanchen in https://github.com/unslothai/unsloth/pull/7454
* Vulkan GPUs: real device names and selectable ordinals (rebase of #7356 onto #7476) by @danielhanchen in https://github.com/unslothai/unsloth/pull/7498
* AMD: enable ROCm torch on gfx906 (MI50 / Radeon VII) on Linux by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7354
* Rotate desktop updater signing key by @wasimysaid in https://github.com/unslothai/unsloth/pull/7500
* Studio: Expose GPU memory mode in unsloth run and unsloth start by @oobabooga in https://github.com/unslothai/unsloth/pull/7421
* Studio: tighten the comments added by the OpenAI model-admission work by @danielhanchen in https://github.com/unslothai/unsloth/pull/7501
* avoid Hub metadata probe when loading tokenizers with local_files_only by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7482
* gitignore: match the stray "~" TMPDIR dir at any depth, plus /temp/ by @danielhanchen in https://github.com/unslothai/unsloth/pull/7499
* tests: stop the installer constraint test counting occurrences by @danielhanchen in https://github.com/unslothai/unsloth/pull/7503
* Installer: opt-in Vulkan llama.cpp backend (and fallback when no AMD card is HIP-supported) by @Souravrajvi0 in https://github.com/unslothai/unsloth/pull/7373
* docs(studio): fix stale gfx110X example in ROCR masking comments by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7440
* Studio UI font-scale test: normalise paths so the allowlists work on Windows by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7434
* Studio voice-tab: use text-ui-* tokens instead of raw px text sizes by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7378
* Studio: split parallel tool calls for Llama 3.x chat templates by @NilayYadav in https://github.com/unslothai/unsloth/pull/7426
* Fix MLX inference Transformers sidecar activation order by @Lyxot in https://github.com/unslothai/unsloth/pull/7402
* Studio: add option to disable the in-memory API monitor by @gaurav0107 in https://github.com/unslothai/unsloth/pull/7156
* Studio: keep automatic model loading toast visible until completion by @oobabooga in https://github.com/unslothai/unsloth/pull/7425
* Surface actionable installer failures in Studio desktop by @oobabooga in https://github.com/unslothai/unsloth/pull/7529
* Studio: detect an interrupted dependency install instead of launching a backend that cannot import by @danielhanchen in https://github.com/unslothai/unsloth/pull/7492
* Default coding-agent servers to reasoning off by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7521
* Fix Claude client tools under server tool policy by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7518
* Handle llama.cpp tool schema limits by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7512
* Isolate Studio PostCSS configuration by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7513
* Fix Windows Codex temporary home path by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7519
* Remove the transient Studio desktop auth handoff by @wasimysaid in https://github.com/unslothai/unsloth/pull/7542
* feat(studio): run chats in parallel in the Chat tab by @danielhanchen in https://github.com/unslothai/unsloth/pull/7455
* Fix image and file paste in Studio desktop by @wasimysaid in https://github.com/unslothai/unsloth/pull/7543
* Studio: preserve durations across reasoning blocks by @oobabooga in https://github.com/unslothai/unsloth/pull/7520
* Studio: Gate fenced-HTML canvas cards on the Canvas toggle by @NilayYadav in https://github.com/unslothai/unsloth/pull/7514
* Studio: match llama.cpp SWA cache sizing by @oobabooga in https://github.com/unslothai/unsloth/pull/7530
* Installer: name the encoding when syncing the prebuilt marker by @danielhanchen in https://github.com/unslothai/unsloth/pull/7554
* Studio tests: bump the tensor-abort mtime by 1ms so the case runs on Windows by @danielhanchen in https://github.com/unslothai/unsloth/pull/7556
* Studio: keep grouped Python scripts visible and save them natively by @oobabooga in https://github.com/unslothai/unsloth/pull/7528
* Pin utf-8 on the two marker reads/writes added with the Vulkan backend by @danielhanchen in https://github.com/unslothai/unsloth/pull/7507
* perf(studio): remove quadratic region scan in LaTeX preprocessing by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7538
* Gate the sed commands that run a shell by @danielhanchen in https://github.com/unslothai/unsloth/pull/7483
* Fix Windows no-torch setup by @Imagineer99 in https://github.com/unslothai/unsloth/pull/7511
* Studio: improve Deep Research synthesis by @oobabooga in https://github.com/unslothai/unsloth/pull/7393
* fix(rocm): prefer system LLVM runtime on native Linux by @Oxygen56 in https://github.com/unslothai/unsloth/pull/7448
* Studio: fix macOS titlebar drag and collapsed layout by @wasimysaid in https://github.com/unslothai/unsloth/pull/7555
* fix(studio): stop re-tokenizing the whole code block on every frame while streaming by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7537
* Studio: run the src-tauri unit tests in CI and fix the two that never ran by @danielhanchen in https://github.com/unslothai/unsloth/pull/7558
* Add Unsloth desktop deep links by @wasimysaid in https://github.com/unslothai/unsloth/pull/7560
* Keep `import unsloth` working when bitsandbytes is absent by @danielhanchen in https://github.com/unslothai/unsloth/pull/7502
* Studio: bound how many tool approvals may park their slot by @danielhanchen in https://github.com/unslothai/unsloth/pull/7496
* Windows setup: route the stale-manifest failure through Exit-SetupFailure by @danielhanchen in https://github.com/unslothai/unsloth/pull/7569
* fix: enable XPU support and update hardcoded CUDA selections for tests by @JoshuaL3000 in https://github.com/unslothai/unsloth/pull/7401
* Studio: pass raise_on_error=False on the stdio MCP call path by @NilayYadav in https://github.com/unslothai/unsloth/pull/7517
* Cover the FP8 row-scaling path in the newer-mapper probe by @vineethsaivs in https://github.com/unslothai/unsloth/pull/7516
* Parse a .json dataset file as one JSON document instead of line-by-line by @vineethsaivs in https://github.com/unslothai/unsloth/pull/7422
* feat(studio): adjustable llama-server parallel slots from the web UI by @knamroud in https://github.com/unslothai/unsloth/pull/7447
* Route the stale-manifest abort through Exit-SetupFailure by @danielhanchen in https://github.com/unslothai/unsloth/pull/7570
* Studio: honour LLAMA_ARG_FLASH_ATTN when recording the launched flash-attention state by @danielhanchen in https://github.com/unslothai/unsloth/pull/7557
* amd: require bitsandbytes>=0.50.0 in the amd extra (fixes ROCm 4-bit NaNs) by @LeoBorcherding in https://github.com/unslothai/unsloth/pull/7535
* fix(studio): show the current artifact's source after switching artifacts by @danielhanchen in https://github.com/unslothai/unsloth/pull/7565
* Studio: surface the tool-call nudge in the chat UI by @danielhanchen in https://github.com/unslothai/unsloth/pull/7559
* Studio: free the llama-server slot when a chat stream reaches [DONE] by @danielhanchen in https://github.com/unslothai/unsloth/pull/7564
* Installer: stop requiring a developer toolchain on the consumer path by @danielhanchen in https://github.com/unslothai/unsloth/pull/7547
* Desktop: ask before quitting on top of a running install by @danielhanchen in https://github.com/unslothai/unsloth/pull/7550
* Tests: import bitsandbytes before the GPU-free harness spoofs CUDA by @danielhanchen in https://github.com/unslothai/unsloth/pull/7582
* Gate the torchcodec audio extras to platforms that have a wheel by @danielhanchen in https://github.com/unslothai/unsloth/pull/7587
* Installer: wrap install.sh in a function so a piped install cannot report curl (56) by @danielhanchen in https://github.com/unslothai/unsloth/pull/7548
* Clear ALLOW_BITSANDBYTES when the bitsandbytes native kernels are not real by @danielhanchen in https://github.com/unslothai/unsloth/pull/7578
* Fix bitsandbytes zombie module breaking test collection on CPU runners by @danielhanchen in https://github.com/unslothai/unsloth/pull/7580
* Show release notes in the update popup, sourced from CHANGELOG.md by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7432
* Studio: read model text as utf-8 so umlauts survive on Windows by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7467
* Fix the red test on main from the #7578 / #7580 merge interaction by @danielhanchen in https://github.com/unslothai/unsloth/pull/7590
* Studio: make the model download folder reachable from the Hub, and findable in search by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7466
* Studio: make the sidebar width draggable by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7561
* Studio: tighten the sidebar pill right inset by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7562
* Measure where Studio's startup time actually goes by @danielhanchen in https://github.com/unslothai/unsloth/pull/7553
* Windows: unblock the consumer install on clean and no-winget machines by @danielhanchen in https://github.com/unslothai/unsloth/pull/7549
* Studio: make the run settings panel draggable by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7566
* Studio: Write auth secret files with a trailing newline by @NilayYadav in https://github.com/unslothai/unsloth/pull/7576
* CI: extract Get-HostMachineArch for the VC++ round-trip test by @danielhanchen in https://github.com/unslothai/unsloth/pull/7597
* Run unsloth_cli/tests in Backend CI by @danielhanchen in https://github.com/unslothai/unsloth/pull/7598
* Make the unsloth_cli studio tests pass in isolation by @danielhanchen in https://github.com/unslothai/unsloth/pull/7599
* Studio: reset-password rotates the credential in place instead of deleting auth.db by @NilayYadav in https://github.com/unslothai/unsloth/pull/7573
* Studio: match the Deep research caret to the other composer pills by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/7601
* Studio: Stop every running Unsloth server, not just the last one recorded by @NilayYadav in https://github.com/unslothai/unsloth/pull/7577
* CI: stop printing the reset passphrase into the agent-guides job logs by @danielhanchen in https://github.com/unslothai/unsloth/pull/7603
* Add test to guard against duplicate keys in __INT_TO_FLOAT_MAPPER by @Suchitra-idu in https://github.com/unslothai/unsloth/pull/7419
* Studio: Don't re-prompt finished answers in the tool loop by @NilayYadav in https://github.com/unslothai/unsloth/pull/7505
* Studio: Drop <header> from fetched page main content by @NilayYadav in https://github.com/unslothai/unsloth/pull/7509
* Extract Get-HostMachineArch for the setup.ps1 Pester suite by @danielhanchen in https://github.com/unslothai/unsloth/pull/7606
* Installer: accept wget in studio/setup.sh, as install.sh already does by @danielhanchen in https://github.com/unslothai/unsloth/pull/7600
* unsloth start: harden coding-agent installation by @oobabooga in https://github.com/unslothai/unsloth/pull/7523
* Studio: use local Unsloth logo for Hub fallbacks by @wasimysaid in https://github.com/unslothai/unsloth/pull/7609
* CI: prove the installer works on a machine with no developer toolchain by @danielhanchen in https://github.com/unslothai/unsloth/pull/7551
* Studio: stabilize floating live monitor geometry by @oobabooga in https://github.com/unslothai/unsloth/pull/7579
* Studio: support Vulkan GPU selection and explicit Vulkan installation by @InfoSage05 in https://github.com/unslothai/unsloth/pull/7188
* Use leading-tight on the truncating sidebar model name by @danielhanchen in https://github.com/unslothai/unsloth/pull/7611
* CI: prove an interrupted install can never masquerade as a healthy one by @danielhanchen in https://github.com/unslothai/unsloth/pull/7552
* feat: allow disabling Studio API monitor logs by @Adityai1411 in https://github.com/unslothai/unsloth/pull/7175
* Export ACCELERATE_MIXED_PRECISION when bf16/fp16 is set explicitly (fixes #4891) by @Sekinal in https://github.com/unslothai/unsloth/pull/7534

## New Contributors
* @Solaris-star made their first contribution in https://github.com/unslothai/unsloth/pull/7286
* @Souravrajvi0 made their first contribution in https://github.com/unslothai/unsloth/pull/7300
* @guerrieroriccardo made their first contribution in https://github.com/unslothai/unsloth/pull/7272
* @lzvxck made their first contribution in https://github.com/unslothai/unsloth/pull/7315
* @JoshuaL3000 made their first contribution in https://github.com/unslothai/unsloth/pull/7136
* @pwasiewi made their first contribution in https://github.com/unslothai/unsloth/pull/7429
* @Oxygen56 made their first contribution in https://github.com/unslothai/unsloth/pull/7448
* @knamroud made their first contribution in https://github.com/unslothai/unsloth/pull/7447
* @Suchitra-idu made their first contribution in https://github.com/unslothai/unsloth/pull/7419
* @Adityai1411 made their first contribution in https://github.com/unslothai/unsloth/pull/7175

**Full Changelog**: https://github.com/unslothai/unsloth/compare/v0.1.50-beta...v0.1.510-beta

## What's Changed in Unsloth-Zoo
* Fix sys.modules leak in offline cross-sync test breaking Core CI by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/926
* Harden Mega-cache artifact loading by @Imagineer99 in https://github.com/unslothai/unsloth-zoo/pull/906
* Fix MLX Metal resource-limit crash in finite text training via CPU batch plans and bounded compile signatures by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/918
* Auto-disable gradient-checkpoint double buffering on unified-memory GPUs by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/928
* Fix Gemma 4 audio and forced-float32 PLE dtype mismatches by @Etherll in https://github.com/unslothai/unsloth-zoo/pull/925
* MLX: faithful fused CCE for scaled/transformed output heads, head-eligibility gating, and label smoothing by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/930
* Fix Gemma-4 vision pooler float16 overflow (26B-A4B / 31B NaN loss) by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/931
* Prefer rocBLAS over hipBLASLt on ROCm RDNA to avoid Composable Kernel GEMM JIT stall by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/921
* fix: unstack hf_to_vllm_mapper for in-memory LoRA on vLLM >= 0.25.0 and enable XPU by @JoshuaL3000 in https://github.com/unslothai/unsloth-zoo/pull/919
* Fix MLX simulation shim children traversal by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/934
* feat(mlx): true grad-norm reporting and compiled global-norm clipping with gradient accumulation by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/922
* Fix CPU MLX shim so output-head resolution tests pass without real MLX by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/935
* Fix allocator fragmentation on Windows/WSL and correct the torch 2.10 alloc conf boundary by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/933
* Fix CPU MLX shim: drop duplicate Module.children/__contains__ by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/938
* Fix Whisper pad token repair by @Etherll in https://github.com/unslothai/unsloth-zoo/pull/936
* feat(mlx): accept bnb_4bit_use_double_quant and cover double-quant loads by @BardiaKoopah in https://github.com/unslothai/unsloth-zoo/pull/911
* fix(mlx): accept Studio VLM singular image batches by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/894
* Preserve MLX-VLM contracts across supported releases by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/898
* Fix MLX LoRA for routed SwitchLinear experts by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/909
* MLX: lazy/streaming dataset support for text and VLM training, length-grouped batching, and background prefetch by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/927
* feat(mlx): CUDA-parity continued pretraining (embed_tokens / lm_head) by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/932
* fix(mlx): freeze base when reloading a LoRA adapter for continued training by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/939
* MLX: bound compiled VLM training signatures with a finite batch plan and shape guard by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/940
* Preserve MLX-VLM compatibility across 0.4.4–0.6.7 by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/952
* Tests: unblock MLX CI on Mac by not requiring the fused MRoPE grad to raise by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/948
* Fix compiler codegen for a forward hidden behind a bare decorator by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/949
* Add HF-style callback support to MLXTrainer by @Lyxot in https://github.com/unslothai/unsloth-zoo/pull/873
* Keep saving_utils importable when bitsandbytes is absent by @danielhanchen in https://github.com/unslothai/unsloth-zoo/pull/953
* hf_utils: unwrap single-adapter peft_config and reject empty model_types by @BardiaKoopah in https://github.com/unslothai/unsloth-zoo/pull/955
* Fix silent embedding corruption in add_new_tokens on padded-vocab mod by @BardiaKoopah in https://github.com/unslothai/unsloth-zoo/pull/923