## unslothai/unsloth — v0.1.702-beta…v0.1.800-beta

_70 commits._

### Features
- **Add an extra llama-server arguments box to the model settings (#8702)** (2098b7c)
- **add a way to open a chat's sandbox folder from the ui (#8661)** (e2b8686)

### Fixes
- **Studio: enforce the research browser regression harnesses (#8736)** (7bf8ade)
- **fix(studio): keep a slow install alive and name what it is downloading (#8805)** (b69bfe0)
- **Studio: fix GGUF image-capability detection for API auto-switch and audio-only projectors (#8807)** (11f59eb)
- **fix(amd): read ROCm VRAM through amd-smi so the backend keeps no HIP context (#8669)** (300292b)
- **fix grouped_gemm crash on the documented gather_indices=None default (#8629)** (57e4bdf)
- **Fix Studio CPU saturation on long streaming replies (#8750)** (d8c1145)
- **Fix Windows HF dataset symlink fallback (#8741)** (13f5bd6)
- **Fix the backend test suite going red on main (#8758)** (4c4ca32)
- **Fix Audio layout and responsive media headers (#8445)** (4b38ab1)

### Backend
- **Bump install.sh / install.ps1 pins to unsloth>=2026.8.17, unsloth-zoo>=2026.8.12 (#8831)** (a8be2a8)
- **Update _version.py** (afa9d81)
- **Update pyproject.toml** (cb7c71c)
- **Offer the media pickers only what the host can run, and name the H3 speed gap (#8822)** (23d670a)
- **Revert "Studio: verify the installed version after the update pass (#8505)" (#8824)** (0fb21e8)
- **Studio: verify the installed version after the update pass (#8505)** (fe79a5b)
- **Revert "Desktop: ship a complete Linux AppImage (#8695)" (#8823)** (83e163f)
- **Studio: apply the provider metadata block to DNS aliases as well (#8782)** (0282789)
- **Hand the LoRA to vLLM on TRL 1.x GRPO rollouts (#8701)** (c3230a7)
- **Studio: stop an unreachable Hugging Face endpoint from stalling the backend (#8799)** (ca7c83f)
- **Studio: never start llama-server at native context on Metal (#8709)** (8bfc61c)
- **Studio: save a chat or reply to project sources (#8588)** (731ea4c)
- **studio: attach youtube transcripts from pasted chat links (#8648)** (6c5385a)
- **Give Unsloth Studio its first CUDA coverage, on a Kaggle T4 (#8489)** (482abfa)
- **Round the aggregate GPU VRAM total so it stops printing float residue (#8820)** (df64e5d)
- **Cover the Audio header in the macOS drag-region contract test (#8718)** (7ae2069)
- **Studio: self-repair sidecars whose extensions were built for another Python (#8705)** (1fb184d)
- **Studio: drag to reorder pinned models in the hub (#8591)** (2899072)
- **Desktop: Keep server stopped across webview reloads (#8567)** (824caca)
- **studio: report prompt speed in tokens per second for safetensors chat (#8647)** (e7c1a13)
- **studio: open linked-folder sources in binary mode on windows (#8621)** (bb67053)
- **Studio: split the Full access rewrite claims per API, and drop the local-desktop promise (#8605)** (19bad46)
- **Wheel the voice model picker once its rows are on screen (#8743)** (ad96b66)
- **Tighten the comments on the versionless torch probe tests (#8811)** (9322559)
- **Exclude studio/backend/tests from the wheel (#8501)** (083d613)
- **Studio: gate --flash-attn, --no-context-shift and --jinja on the binary (#8710)** (70a0f1a)
- **studio: give minimax-h3 the sd.cpp graph-cut flags so it renders on a smaller card (#8644)** (753ed31)
- **Carry a provider-run tool's result into the next turn (#8713)** (02f63cd)
- **Studio: run the Transformers dictation engine in a spawn child (#8668)** (62b12f0)
- **Studio: optimize startup with on-demand language catalogs (#8623)** (eb869b8)
- **Studio tests: pin the empty-version torch probe distinction at the repair paths (#8803)** (a0c1723)
- **Ask for the transformers a training model needs, before the run starts (#8721)** (d50f49e)
- **Read the descender guard's class list wherever it is written (#8765)** (8fef5ed)
- **Studio: budget the Metal context fit from free memory, not total (#8706)** (9cc6e65)
- **Make two backend test guards able to fail (#8760)** (8c6048a)
- **Studio: run the LoRA scan off the event loop (#8392)** (71ad2b3)
- **Run deterministic notebook smoke tests on real Kaggle T4s (#8440)** (bb11b37)
- **Studio tests: pin the torch probe invalidation contract to both installers (#8795)** (cb07e02)
- **Windows: stop compiling C# for colour on hosts that already render it (#8767)** (9c69529)
- **Studio: close seven detection gaps in the deliberate-crash guard (#8788)** (eb36701)
- **Desktop: ship a complete Linux AppImage (#8695)** (93a77d4)
- **Studio: show live prompt and generation speed (#8700)** (ea22195)
- **Studio: never signal pid 0 or pid 1 from the lifetime reaper (#8797)** (7e3039e)
- **Studio tests: do not dump core when a probe child is crashed on purpose (#8783)** (1838e10)
- **Credit the contributor PRs behind the external-provider tool loop (#8704)** (b16a40a)
- **Studio: share one torch classification probe across the repair paths (#8779)** (7fd3eeb)
- **Let a request say whether Search runs locally or at the provider (#8711)** (015b4a0)
- **Studio: keep the RAG embedder off the GPU unless asked (#8654)** (34a1197)
- **Studio: tighten the Settings > Logs comments (#8769)** (ea34ac9)
- **Studio: read the logs from inside the app, in Settings > Debugging (#8690)** (046e3ee)
- **Use Kaggle's large overlay for saves, and refuse a GGUF export that cannot fit (#8439)** (48ea556)
- **Studio: idle auto-unload for the image and video backends (#8672)** (32c2627)
- **studio: carry chat settings across browsers and remote sessions (#8656)** (c38daa3)
- **Stub the trainer's heavy imports in the audio-probe test so Backend CI collects again (#8740)** (5ebcc4f)
- **Let the uv signal handler exit without going through setup_fail (#8744)** (e227792)
- **Bump install.sh / install.ps1 pin to unsloth>=2026.8.16 (#8714)** (10f34db)
- **Update mapper.py** (41ef68d)
- **Studio: read a model picker row at a glance (#8658)** (8d0edc1)

### Chore
- **CI: collect the zoo sanity suite from the zoo checkout (#8737)** (44afbdf)

_Recap by [Repo Wrapped](https://repowrapped.com/gh/unslothai/unsloth?utm_source=github-action)._