unslothai/unsloth v0.1.913-beta on GitHub

Make Unsloth your own with custom backgrounds and a improved UI/UX. This release also cuts app update times and speeds up Linux app and Qwen-Image gen when VRAM is limited. Also new: DeepSeek-V4.1-Flash-1017 in Unsloth Dynamic v3.1 GGUFs (preview), FLUX.2-klein-9B on low-RAM machines, smoother chats and batched MoE fine-tuning.

Highlights

Custom backgrounds to personalize your chats (more customization options coming next week)
DeepSeek-V4.1-Flash-1017 new Unsloth Dynamic v3.1 GGUFs (preview), the most accurate for their size on nearly all benchmarks Blog
14 new languages for Unsloth
App updates finish twice as fast
Linux much less lag and improved performance
Qwen-Image renders up to 2x faster under a 16 GB VRAM limit
Updated interface with consistency and cleaner UI
MoE fine-tuning runs compatible API jobs together

Customization and interface

Move the chat box to the bottom of new chats with Composer at the bottom in Settings > Chat.
Hide the greeting entirely for a simpler new-chat screen.
The Library gets shorter chat cards, clearer tooltips and consistent sort and filter icons.
Copy images between input slots by dragging a filled slot onto an empty one.
Pinned models stay saved across Unsloth Desktop restarts.
Each browser tab keeps its own unsent message.
Scrolling long chats is smoother on Linux with NVIDIA, X11 and driver 610+. Our high-resolution test improved from about 17 to 60 fps.

Images and video

FLUX.2-klein-9B now loads and renders on Linux with 12.7 GB of system RAM and a supported GPU.
Qwen-Image-2512 renders up to 2x faster under a 16 GB VRAM limit.
Qwen-Image-Edit-2511 renders about 1.3x faster under the same limit.
MiniMax-H3 clips finish about 1.7x faster in our memory-limited NVIDIA test.

Performance

App updates take about half as long by skipping Python precompilation. The tradeoff is a longer first launch afterwards, as before v0.1.912.
Long replies stream more smoothly, even with unfinished math formatting. Parsing an 83 KB reply with a stray $$ took 0.7 seconds instead of 71 seconds.
Previewing a 1 GB local Parquet file uses about 300 MB of peak memory, down from 1.5 GB.
Unsloth stays responsive as Transformers and MLX models finish loading.

Chat and API

GGUF models reload in the background after a llama-server crash.
Failed or cancelled GGUF downloads leave your current model loaded.
Very long GGUF chats can compact and keep going.
Claude Code connected directly to a local GGUF reuses cached prompts across new sessions.
Rejected GGUF requests show llama-server’s explanation instead of a generic error.

Mac

Claude Code, Codex and OpenCode requests can share an MLX batch. In our test, four simultaneous requests finished up to 2.3x faster.
Experimental Neural Engine Prefill processes prompts up to 31% faster in our Qwen3.8-27B-4bit test on M3 Max. It is opt-in and can slightly change replies.
Auto context starts GGUF models at no more than 65,536 tokens, reducing upfront memory use. You can still set a larger context manually.
Send LFM3-Audio or LFM3-Omni a voice clip on MLX and hear its answer alongside the text.

Training

Compatible LoRA and QLoRA jobs on MoE models such as Qwen3.6-35B-A3B train together through the fine-tuning API, each with its own adapter.
Use via API on the Train page opens Python, JavaScript and curl examples for your setup.
Single-column text datasets are recognised automatically, with no manual column mapping.

What's Changed

Bump install.sh / install.ps1 pins to unsloth>=2026.10.10, unsloth-zoo>=2026.10.10 by @danielhanchen in #15785
Tensor parallel x FSDP2: average trainable TP parameters left out of FSDP2 over the data parallel ranks by @danielhanchen in #15748
Flash impl: warn about an explicit fa3 / fa4 fallback on the Trainer path too by @danielhanchen in #15750
FSDP2 block CUDA graphs: free the gathered weights of flex BlockMask blocks by @danielhanchen in #15756
Installer downloads: report progress on whole 1 MiB chunks again by @danielhanchen in #15789
Studio export: skip the GPU cache clear when nothing was loaded since the last one by @danielhanchen in #15737
Studio video: read an LTX-2 checkpoint header once per file when probing for 2.3 by @danielhanchen in #15738
Tensor parallel: batch_rebalance and group_by_length samplers give tp peers the same rows by @danielhanchen in #15752
CI: hold the unpinned version-compat lanes below torch 2.15 until torchvision 0.30 is published by @danielhanchen in #15774
Studio: give each batched fine-tuning job its own fp16 loss scale by @danielhanchen in #15741
Studio: list a Qwen-Image-2.1-Turbo base pipeline cached in its FP8 mirror On Device by @danielhanchen in #15772
Multi-LoRA: per-adapter LoRA on fused MoE experts by @danielhanchen in #15742
Frontend tests: follow the desktop quit hiding the window by @danielhanchen in #15801
Cross entropy: 8192-wide chunks at 4 warps for the chunked forward on sm120 by @danielhanchen in #15744
Run packed rows of head-dim-routed flex decoders through varlen SDPA by @danielhanchen in #15514
Tests: allow the zoo's padded logits row stride in the fused linear adapter check by @danielhanchen in #15803
studiobench: seed the browser with update notifications off by @danielhanchen in #15805
Flex small-smem tiles: keep float32 Gemma 3 off flex on sm86 / sm89, tile forced no-grad calls by @danielhanchen in #15529
SDPA decode: keep the caller's backend policy when avoiding cuDNN by @danielhanchen in #15525
MoE heuristic dW config: 128-wide N tile for gpt-oss K = 2880 by @danielhanchen in #15505
FP8: skip the compiled wrapper for rowwise FP8 inference linears by @danielhanchen in #15584
Studio: size FLUX.1 Kontext guard, preview and compile shape from its ~1 MP render by @danielhanchen in #15797
Keep image and video models loaded beside chat models on Apple Silicon by @Lyxot in #15693
Block CUDA graphs: turn on automatically for fixed-shape, launch-bound LoRA runs by @danielhanchen in #15768
Studio: batch MoE fine-tuning jobs with per-job expert LoRA by @danielhanchen in #15809
Studio: raise the startup bundle budget by the Trick or treat row's growth by @danielhanchen in #15822
Studio: test that a batched fine-tune stays in its own account by @danielhanchen in #15825
Studio: keep MiniMax-H3's streamed top-level group on the card between denoise steps by @danielhanchen in #15796
Studio native cache write: recapture when an offload moves weights before the live-read swap by @danielhanchen in #15776
checkpoint_skip_layers="auto": skip the planning collective under DDP.join() by @danielhanchen in #15779
FP8 scaled_mm backend: send 17-64 rows to torch._scaled_mm when the 16-bit weight spills L2 (sm89) by @danielhanchen in #15791
CI: turn update checks off in every job that drives a browser by @danielhanchen in #15815
Studio tests: drain the background sandbox mover between tests by @danielhanchen in #15830
Studio: emit no legacy load flag when the llama-server --help probe is inconclusive by @danielhanchen in #15828
Mamba LoRA: apply a trained dt_proj / x_proj / out_proj adapter in eval and cached decode by @danielhanchen in #15730
Studio diffusion: load the int8 GEMM operands through TMA descriptors on sm120 by @danielhanchen in #15829
Block CUDA graphs: only release the cudagraph pool for tree functions the block backend compiled by @danielhanchen in #15777
Tests: FP8 routing tests pin the NVIDIA path so they pass on ROCm runners by @danielhanchen in #15838
Studio: skip the Keep resident page-lock when the llama-server --help probe is inconclusive by @danielhanchen in #15833
Backend CI: split the fsdp2 Repo tests shard in two by @danielhanchen in #15851
Fix OpenCode Go session headers in Studio by @Aafimalek in #15780
Desktop: re-enable GPU compositing on NVIDIA X11 with driver 610 or newer by @oobabooga in #15546
Studio tests: skip rag_conn where this Python's sqlite3 cannot load extensions by @danielhanchen in #15862
GRPO: take the per-step host syncs out of the loss path by @danielhanchen in #15855
Studio: keep refusing a symlinked spill where there is no O_NOFOLLOW (Windows) by @danielhanchen in #15887
Studio: show the real error when a Responses stream fails by @NilayYadav in #15816
Studio: show why llama-server rejected a chat request by @NilayYadav in #15817
Packed attention: varlen probes under fullgraph + dynamic=True compile by @danielhanchen in #15891
Studio tests: keep GGUF loads off the host's Metal budget unless a test asks by @danielhanchen in #15910
Studio tests: measure the int8 / fp8 dequant and nibble-swap peaks in a fresh process by @danielhanchen in #15912
Studio: restore the #15731 login layout by @shimmyshimmer in #15903
Studio: lead the Live monitor tiles with the reading, unrounded by @shimmyshimmer in #15897
Studio: put a just-created API token in the Agents command and API examples by @shimmyshimmer in #15890
Studio: Images icon for Reference, file search icon for Chat with files by @shimmyshimmer in #15806
Studio: Qwen-Image-2.1 eager renders before the deferred compile now match ComfyUI by @danielhanchen in #15908
Studio diffusion: let a 12.7 GB host load FLUX.2-klein-9B by spilling the int8 denoiser to a scratch file by @danielhanchen in #15911
Studio: stop the first Wan2.2-A14B render from waiting for the second expert's pin by @danielhanchen in #15906
Studio tests: give managed engines ports another worker cannot be handed by @danielhanchen in #15915
Studio: add a Quickstart heading over the API example tabs by @shimmyshimmer in #15898
Studio: one text and control width for every settings row by @shimmyshimmer in #15905
Studio: Decision tab loads at once, System and settings polish, main-theme borders restored by @shimmyshimmer in #15919
Studio: decode format-check preview rows once per request by @danielhanchen in #15867
Studio: index the training history's resumed-later check by @danielhanchen in #15872
Studio: copy Whisper timestamp tokens to the host once per window by @danielhanchen in #15860
Studio: fetch up to four Hub files at once over HTTP when partials resume by @danielhanchen in #15781
Studio: reuse the VLM URL probe's downloads in the conversion by @danielhanchen in #15870
Studio: fetch ShareGPT+image row images concurrently, in row order by @danielhanchen in #15786
Studio desktop: ask desktop-capabilities in the launch probe on a capability cache miss by @danielhanchen in #15823
Studio: find a code block's backtick runs by regex when picking its fence by @danielhanchen in #15913
GRPO PrefixGrouper: import cleanly on torch 2.6 and keep it off below 2.7 with a log line by @danielhanchen in #15836
Studio: write transcoded WAV blocks raw and fix the header once on close by @danielhanchen in #15917
Studio: find VLM rows without an image or caption from Arrow validity by @danielhanchen in #15787
Studio: score unclosed Markdown link destinations in linear time by @danielhanchen in #15811
Studio: write each training metrics flush and its run progress in one transaction by @danielhanchen in #15871
GRPO PrefixGrouper: build the shared-prefix layout from one host copy instead of per-token loops and per-segment device fills by @danielhanchen in #15835
Studio: reuse recipe shard row counts while no shard changed by @danielhanchen in #15869
Studio: make Recipes lint validators run on Windows and under a gitignored temp root by @danielhanchen in #15929
Studio: wake admission waits on queue state changes instead of their next poll by @danielhanchen in #15885
Studio: treat unknown parent liveness as alive in the llama-server reaper by @danielhanchen in #15921
Studio: stop re-hashing kept spills and re-reading the record on every prune by @danielhanchen in #15884
Studio: adopt the WAV save_from_media already transcoded instead of decoding it again by @danielhanchen in #15916
Cleanup: remove 5 redundant pass statements that leave bytecode unchanged by @danielhanchen in #15790
Studio setup.ps1: install bun only when the frontend is built by @danielhanchen in #15933
Studio: keep the Markdown block split linear after an unmatched $$ by @danielhanchen in #15886
Tokenizer loading: skip the unused slow tokenizer load when the fast one is returned as is by @danielhanchen in #15834
Import fixes: disable a torchao too new for the installed torch instead of failing import unsloth by @danielhanchen in #15850
Studio: make the install-time bytecode precompile opt-in by @danielhanchen in #15923
Studio: read the file digest cache once per reuse estimate by @danielhanchen in #15914
GGUF export: write the intermediate conversion inside the output folder so placing it is a rename by @danielhanchen in #15783
Studio: build the MCP apps on first use instead of at backend import by @danielhanchen in #15814
Studio: read only the run state for chunk-only chat run event batches by @danielhanchen in #15807
Studio: skip the article and main scope passes when the page has no such tag by @danielhanchen in #15812
Studio: preview local Parquet from its leading rows instead of loading it whole by @danielhanchen in #15868
Studio: store float32 RAG vectors from their buffer and pack the query once by @danielhanchen in #15859
Studio: share one in-flight /api/health read between the startup and route platform checks by @danielhanchen in #15794
Studio tests: use a portable fake llama-server in the load cancel tests by @danielhanchen in #15826
install-kernels: skip the torch import when a kernel's extension is absent, and probe torch only when a wheel is needed by @danielhanchen in #15896
Studio: validate a recipe job's OXC batches in one Node process by @danielhanchen in #15925
Studio: walk the sandbox once per tool call boundary for the files card and file history by @danielhanchen in #15793
Studio: keep audio gallery operations in the folder they resolved by @danielhanchen in #15875
Studio: skip the escaped-math parse when no escaped dollar pair can be math by @danielhanchen in #15808
Studio: quantise an offloaded native int8 denoiser on the card, one linear at a time by @danielhanchen in #15936
GRPO PrefixGrouper: fit the flex kernel on 101 KB shared-memory GPUs and stop its silent eager fallback by @danielhanchen in #15854
Studio: skip the route's GPU probe when the GGUF fits in RAM by @danielhanchen in #15902
Studio desktop: skip the device type read while the backend port is still unknown by @danielhanchen in #15824
Studio Desktop: sync pinned models and overrides after the backend port is known by @danielhanchen in #15821
Studio: matching Hugeicons permission icons, larger and top-aligned by @shimmyshimmer in #15944
Tests: baseline the torchao import probe from #15850, gate the PrefixGrouper CUDA test on a real card by @danielhanchen in #15948
Tests: follow the media header's MediaHeaderActions row from #15919 by @danielhanchen in #15946
Studio: accent update buttons, standard switches, one icon per meaning by @shimmyshimmer in #15939
Studio: greeting setting hides the whole greeting, composer at the bottom option, no model disclaimer by @shimmyshimmer in #15904
Studio: write plain recipe export rows with one json.dumps call by @danielhanchen in #15926
Studio: reuse the cached TLS context for llama-server health probes by @danielhanchen in #15792
Studio: read the run history once per checkpoint listing by @danielhanchen in #15873
Studio: keep an appearance edit made while the personalization load is in flight by @danielhanchen in #15953
Studio diffusion: run both guidance branches per streamed block upload for Qwen-Image-Edit under a VRAM budget by @danielhanchen in #15930
Tests: read _gpu_init.py as UTF-8 in the torchao-too-new test by @danielhanchen in #15957
Studio: encode a batch's gallery PNGs concurrently by @danielhanchen in #15909
Studio: skip the audio gallery prune's sidecar pass under both caps and list prepared inputs once per sweep by @danielhanchen in #15874
Tests: mac tab walk waits for the sidebar to mount before judging a pinned row by @danielhanchen in #15959
Studio video: MiniMax-H3 conditioner pin yields to a streamed denoiser on a small host by @danielhanchen in #15955
Studio video: MiniMax-H3 unfused Q / K / V share one ConvRot rotation and act quant by @danielhanchen in #15956
CI: split Repo tests (CPU, rest) and (CPU, gloo) before they reach their 2100 s cap by @danielhanchen in #15951
Studio: say when remote access Start automatically opens a public URL by @NilayYadav in #15849
Studio: encode each recall query once per compaction and let archive widening widen by @danielhanchen in #15960
Studio: answer dense RAG search from an in-memory numpy matrix kept in sync by SQLite triggers by @danielhanchen in #15961
Studio: match linked-folder paths against .gitignore runs, not rule by rule by @danielhanchen in #15949
Studio: link to image training from the Train page by @NilayYadav in #15391
Studio diffusion: a first load downloads the weights before the small-host decision reads them by @danielhanchen in #15974
Let decision models export to GGUF on a fresh install by @NilayYadav in #15864
Studio: stop Auto context from reserving a model's whole window on a Mac by @NilayYadav in #15843
Tests: keep tqdm's monitor thread out of the ShareGPT no-threads count by @danielhanchen in #15992
Studio: keep the user's git identity and gh login in unsloth start pi by @NilayYadav in #15865
Studio: let unsloth start vibe start with smaller context windows by @NilayYadav in #15863
Studio: restart a crashed GGUF model instead of showing it loaded by @NilayYadav in #15847
Tests: make unsloth_cli/tests/test_start.py pass on Windows by @danielhanchen in #15987
Studio: keep the loaded model if a GGUF download fails or is cancelled by @NilayYadav in #15846
Studio: test that escaped currency keeps long replies streaming fast by @NilayYadav in #15820
Studio: stop re-uploading the whole chat while a reply streams by @NilayYadav in #15818
Studio: nvfp4 safetensors twins: converter 0-dim compare, refuse a twin the installed torchao cannot unflatten by @danielhanchen in #15977
Studio: keep the app responsive as an MLX or Transformers load finishes by @NilayYadav in #15842
Studio: stop the KV cache type following the last model to the next one by @NilayYadav in #15819
Studio: share one edge index for Recipe Studio node connection status by @danielhanchen in #15942
Studio: load downloaded models with fewer Hugging Face requests by @NilayYadav in #15844
Tests: plant uv.exe on Windows in the MLX repair minimal-PATH lookup test by @danielhanchen in #15995
Studio: reuse lexical passes and branch checks across a compaction and prune branch-run starts by @danielhanchen in #15972
Studio: recommend a Hub GGUF quant that fits the Mac's Metal budget by @NilayYadav in #15848
Studio: stop the Claude CLI billing line from breaking prompt caching by @NilayYadav in #15845
Studio video: stage an fp32 encoder on the card when its dense cast would not fit host RAM by @danielhanchen in #15975
Studio setup: skip the OXC validator npm ci when its installed tree is unchanged by @danielhanchen in #15947
Studio diffusion: use an installed HIP SDK for the Windows ROCm sd.cpp build when its bin dir is not on PATH by @oobabooga in #15853
Studio: train single-string-column datasets as raw text by @YakshithK in #15899
Unsloth Studio: LFM3-Audio / LFM3-Omni on MLX, with spoken replies by @Lyxot in #15889
fix(studio): attach an explicitly chosen MLX drafter even when it is priced not to fit by @Lyxot in #15813
Tests: keep the RAG query-packing test on vec0 while the in-memory cache builds in the background by @danielhanchen in #16013
Studio: load the chat page on first use instead of in the startup bundle by @danielhanchen in #15943
Move the Unsloth Studio MLX pins to mlx-vlm 0.7.11 by @Lyxot in #15861
Tests: price the auto-context fit floor as the layer-split path on any host by @danielhanchen in #16014
Collapse blank line runs with a literal-led pattern (about 10x faster) by @danielhanchen in #15950
Studio diffusion: Qwen-Image text-to-image activation seeds keep part of the int8 DiT resident under a VRAM budget by @danielhanchen in #15985
Studio video: request a hosted .safetensors twin before each declared prequant .pt by @danielhanchen in #15980
Studio: batch MLX tool-protocol, response_format and video replies; several clips and Gemma 4 video on MLX by @Lyxot in #15920
Studio: Use via API button on the Train page by @danielhanchen in #15996
Studio: say why AI Assist cannot map a dataset with no prompt/response split by @danielhanchen in #15991
T4: hand batch-4 NF4 decode to the bitsandbytes gemm_4bit where it runs by @danielhanchen in #15988
Tests: read the training API job bodies fixture as UTF-8 by @danielhanchen in #16025
Studio: hand the audio route's heal gate the request's contract, and follow the hoisted audio_base64 in the turn-scope check by @danielhanchen in #16023
Studio: Neural Engine Prefill toggle for MLX models by @Lyxot in #15954
Studio video: compare streamed frames without chunk-sized temporaries (H3 export stalls) by @danielhanchen in #16024
Studio video: streamed mp4 byte-identical to the stock export on every render (cap libx264 at AVX2 on AVX-512 hosts) by @danielhanchen in #16030
Studio: int8 Qwen-Image text stream when the denoiser is streamed (1.24x under a 16 GiB budget) by @danielhanchen in #16027
Studio MCP: bound get_job wait_s with a fixed limit, not the patchable MAX_WAIT_S by @danielhanchen in #16033
Batched MoE LoRA: every job's expert LoRA in one grouped pass on Zoo's kernels by @danielhanchen in #16012
Studio: toast icons centred on one line, top-aligned on longer toasts by @shimmyshimmer in #16050
Studio: one set of sort and filter icons, shorter chat cards, toolbar tooltips by @shimmyshimmer in #15958
Desktop: Add GPU rendering setting with a Linux launch-crash fallback by @Etherll in #15482
GRPO: keep masking completions cut by the vLLM context limit by @danielhanchen in #16057
Packed mRoPE flash attention: cover a padded tail, same decision when compiled by @danielhanchen in #16058
Cross entropy: trace under torch.compile(dynamic=True) by @danielhanchen in #16059
Scope WavLM's 4-bit attention skip to WavLM's own layers by @danielhanchen in #16060
gpt-oss: accept an absolute transformers_weights inside the local model dir by @danielhanchen in #16061
MXFP4 torch decode fallback: run ldexp on the operands' GPU by @danielhanchen in #16062
FP8 scaled_mm: keep native fp16 output on torch >= 2.14 by @danielhanchen in #16063
Context parallelism: refuse it beside tensor parallelism instead of training the wrong loss by @danielhanchen in #16068
Studio video: batch the CFG pair into one call on a compiled resident Wan denoiser by @danielhanchen in #16067
Context parallel: allow text-only Qwen3.5 checkpoints that carry a vision_config by @danielhanchen in #16072
Studio: keep the composer still when switching chats in WebKit by @shimmyshimmer in #16076
Studio: stop the chat composer draft mirroring into other open tabs by @debojyoti-dev in #15795
float16: agree on GradScaler skips across pipeline stages, scale the CP start, probe the float16 FP8 GEMM by @danielhanchen in #16070
Studio: scale the token count timeout so long chats can compact by @oobabooga in #15900
fix(studio): list approved custom-code models in /v1/models by @Lyxot in #15945
Studio: drag an image from one slot to another by @NilayYadav in #15876
Studio: name a too-short context length when completion masking drops rows by @danielhanchen in #16000
Studio: align the Images and Video preset dropdown with the title's top by @shimmyshimmer in #16081
Tests: stub @/lib/folder-icons in the Hub GGUF card Metal budget test by @danielhanchen in #16082
Studio: fill locale gaps and translate the sign-in, consent and Desktop startup screens by @shimmyshimmer in #16069
Studio: no text selection from the clicks that follow a dropdown dismissal by @shimmyshimmer in #16077
Studio: let conversation search return every earlier turn that fits by @NilayYadav in #16036
Desktop: run the AppImage on the host's WebKitGTK 2.54+ when available by @Etherll in #15503
Studio: add Spanish (Latin America) as a display language by @shimmyshimmer in #16089
Studio: include Schema transform output in recipe dataset downloads by @NilayYadav in #16039
Studio: strip only harmony protocol tokens from gpt-oss answers by @danielhanchen in #16004
Studio chat: publish held stream text on the next frame instead of waiting for the next chunk by @danielhanchen in #15998
Studio: stop skipping scheduled runs right after Studio starts by @NilayYadav in #16037
Studio: batch the raw text blank row filter and EOS append for map-style datasets by @danielhanchen in #15978
Studio: read a JSON body inside <function=...> instead of running the call with no arguments by @danielhanchen in #16008
Tests: run the MXFP4 operand device guard check on CPU-only torch too by @danielhanchen in #16088
Report measured model loading and 4-bit conversion during image training by @Imagineer99 in #15536
Studio: let managed accounts use local MCP servers when allowed by @NilayYadav in #16044
Studio: int8 Qwen-Image streamed modulations through a fused GEMV (1.23x under a 16 GiB budget) by @danielhanchen in #16087
Studio: size a cached DFlash sidecar from disk instead of asking the Hub on every load by @danielhanchen in #15986
Bump trufflesecurity/trufflehog in / by @dependabot[bot] in #15968
fix(studio): keep a long MLX vision chat's prompt snapshot past the cache budget by @Lyxot in #16079
Bump swatinem/rust-cache in / by @dependabot[bot] in #15966
Bump github/codeql-action init and analyze to v4.39.1 by @dependabot[bot] in #15964
Bump actions/setup-python in / by @dependabot[bot] in #15963
Bump the npm-oxc-validator group in /studio/backend/core/data_recipe/oxc-validator with 3 updates by @dependabot[bot] in #15962
Studio: keep a resident GGUF's repo identity for Hub shared-blob cache files by @danielhanchen in #16086
fix(eval): reserve explicit YAML task names by @hsusul in #15984
Studio: offer video input only when llama-server can launch ffmpeg and ffprobe by @danielhanchen in #16094
Tests: follow the localized update banners from #16069 by @danielhanchen in #16090
Studio: keep attached documents in Deep Research reports by @NilayYadav in #16019
Desktop: switch the theme without a view transition on Linux by @oobabooga in #16040
Studio: recover Deep Research from provider JSON validation failures by @alkinun in #15788
Studio: ask llama-server again with its tool parser off when it rejects a reply before streaming by @danielhanchen in #16002
Studio: default Deep Research to no time limit per model request by @NilayYadav in #15983
Studio installer (Windows): avoid unavailable CUDA torchao wheels by @oobabooga in #16016
Studio: share one in-flight media model index build so the startup checks do not scan twice by @danielhanchen in #16075
Add custom hourly intervals for scheduled tasks by @Imagineer99 in #15799
Studio Images: enable GGUF LoRAs on MLX and diffusers, and explain when unavailable by @oobabooga in #16053
Studio: read streamed dataset slices in batches instead of row by row by @danielhanchen in #16083
Studio diffusion: test-launch sd.cpp binaries without the Windows missing-DLL box by @oobabooga in #16031
Studio diffusion: Qwen-Image text-to-image streamed placement uses its own activation figure under a VRAM budget by @danielhanchen in #16095
Studio: download the commit the manifest was planned from by @danielhanchen in #16047
Studio: vector cache in float32 unit rows with streamed builds, in-place appends and bounded block scans by @danielhanchen in #16093
Studio: add customizable chat background images and effects by @wasimysaid in #15973
Studio: keep a bare-JSON tool call out of the final answer after the tool loop by @danielhanchen in #16003
Studio: add OpenCode Zen and OpenCode Go to Connections by @NilayYadav in #16010
Studio: honour speed on /v1/audio/speech by @NilayYadav in #16020
Studio: compile the SDXL U-Net per block so the 4th image no longer waits for the whole compile by @danielhanchen in #16085
Studio: stop offering the VibeVoice q4-ios GGUF that audio.cpp cannot load by @oobabooga in #16048
Studio: recover a closed tool call whose single code argument has unescaped quotes by @danielhanchen in #16006
Studio: add MiniMax to Connections by @NilayYadav in #16011

New Contributors

@YakshithK made their first contribution in #15899
@hsusul made their first contribution in #15984

Full Changelog: v0.1.912-beta...v0.1.913-beta

unslothai/unsloth v0.1.913-beta Customize + Performance + UI/UX improvements on GitHub