unslothai/unsloth v0.1.914-beta on GitHub

Unsloth now has Glass themes and more customization options for backgrounds, greetings and chats. This release also brings faster generation, more multi-GPU training options and CPU training in Unsloth.

Highlights

New Glass themes with 5 themes, transparency + native macOS glass
Customization with custom greetings, background controls and code fonts
DeepSeek-V4.1 512K context benchmarks show Unsloth's GGUFs rank strongest across Terminal Bench, DeepSWE and AutomationBench. Analysis
Chat navigation to preview, bookmark and jump between earlier turns
CPU text training directly from the Train page

Make it yours

Choose from five Glass styles with adjustable strength and optional glass menus, with or without a background image.
Unsloth Desktop on Mac supports native glass, letting your desktop show through the window.
Adjust background size and position, and choose where it appears across your chats.
Add your own greeting with text, an icon or an image, or hide it entirely.
Customize code fonts, line numbers, word wrapping and ligatures with a live preview.
Use the Trick or Treat theme year-round without the seasonal extras.

Chat

Preview earlier messages, bookmark turns and jump through long conversations with the new chat navigator.
Paste formatted content as Markdown, including links, lists and spreadsheet tables.
Arabic, Persian, Hebrew and Urdu messages display right to left while code keeps its original direction.
Long chats stay more responsive while replies stream and you scroll.
MCP connections recover after a server restart.
Coding agents can follow instructions from files in your project's working folder.

Performance and media

Supported single-GPU Llama and Qwen3 models generate faster in Unsloth's Transformers backend on NVIDIA GPUs with Linux or WSL. In our A100 test, 4-bit Qwen3-8B reached 99.1 tokens/s, up from 25.7, a 3.9x speedup.
SDXL now renders under a 4 GiB VRAM budget, tested on T4 and L4 GPUs.
Search large document collections faster with a new index for RAG.
Save generated images as PNG faster without changing the output.
Video models reuse cached INT8 weights for faster repeat loads.
Wan2.2 A14B uses less system memory when building INT8 weights for offloaded generation.
Adjust image LoRA strengths without reloading adapter files.
Transcribe audio up to four hours long with audio.cpp, subject to upload and storage limits.

Mac

Try experimental Fast Verify for faster generation with supported MLX speculative decoding models. Replies may differ slightly.
MLX streaming spends less time processing text as replies grow.
Unsloth now uses the Mac system font by default.

Training

Fine-tune text models on CPU from the Train page when no GPU is available.
Chat-only installs can add CPU training through Install Training.
Fine-tune audio vision models such as Gemma 3n and Gemma 4 using audio and images together in the same dataset.
Load supported non-quantized models larger than a single GPU by loading only each GPU's assigned pipeline layers.
Configure supported combinations of data, tensor, pipeline and context parallelism with UNSLOTH_PARALLEL.
Reduce MLP activation memory during training with unsloth.tile_mlp(model).
Send images to vision models through CLI chat and inference commands.

What's Changed

Bump install.sh / install.ps1 pins to unsloth>=2026.10.11 by @danielhanchen in #16131
Don't warn about loss_type = "nll" on a TRL whose SFTConfig predates it by @debojyoti-dev in #16042
Studio MLX estimate: ask the quantized-attention route about a cache longer than the chunk by @debojyoti-dev in #15970
Tests: the supplied-component twins check runs where diffusers is not installed by @danielhanchen in #16126
Tests: the titlebar smoke treats Vite outdated pre-bundle 504 as a reload by @danielhanchen in #16130
Flex wrapper: send short no-grad queries to the main kernel where flex_decoding cannot compile by @danielhanchen in #16092
Studio Windows AMD: pair triton-windows with torch after the dependency stack by @danielhanchen in #16074
Studio: warn when a Windows GPU load leaves the system short of commit by @debojyoti-dev in #16046
Studio: load LFM3 custom-code models (route by saving transformers version, fix generate, chat and export) by @danielhanchen in #16096
Tests: fused LoRA TP parity compares O and dX at each tensor fp32 scale by @danielhanchen in #16144
Studio: download files over 50GB from ModelScope over HTTPS (#15989) by @indrajeetapache in #15994
Dependencies: pin pymupdf 1.28.2 on Windows ARM64, the only release with a win_arm64 wheel by @danielhanchen in #16160
Studio: MOSS-Image negative-branch layer skip follows the reference workflows per checkpoint by @danielhanchen in #16127
Studio MLX estimate: price attention scores where mlx really holds them, at the width it holds them by @danielhanchen in #16021
Studio: HNSW (usearch) tier for gigantic RAG scopes, kept consistent with SQLite by @danielhanchen in #16145
Tests: the skipped-menu linearity check times each size as the best of five by @danielhanchen in #16170
Dependencies: the studio extra carries usearch like studio.txt by @danielhanchen in #16176
Studio Images: SDXL-Turbo opens at its native 512x512 canvas by @danielhanchen in #16168
Tiled MLP: one flag or one line for any model, on the new sequence tiling engine by @danielhanchen in #16125
Tests: follow the tiled MLP hooks from #16125 (encoding, loader stubs, trainer logger) by @danielhanchen in #16196
Vision GC offload: each model gets its own non-reentrant offloader by @danielhanchen in #16194
LoRA: add the rank-r term inside the base GEMM so small updates are not rounded away by @danielhanchen in #16071
Tests: compare the disabled 4-bit fused LoRA forward with dequantize + linear, not bitsandbytes' fused GEMM by @danielhanchen in #16252
Pipeline parallel: every torch.distributed.pipelining schedule (Interleaved1F1B, LoopedBFS, zero bubble, ZBV, DualPipeV) by @danielhanchen in #15655
Give the WMI deadline check headroom for Windows PowerShell job startup by @danielhanchen in #16257
Run the dynamic-rank LoRA compile test on hosts without Triton by @danielhanchen in #16256
Keep tensor subclass types through the fused LoRA accumulates by @danielhanchen in #16249
Pipeline parallel: per-stage sharded loading (load models larger than one GPU) by @danielhanchen in #15656
Tensor parallel: support few KV heads, full finetuning, weight-only QAT, eval, MoE and dropout instead of refusing by @danielhanchen in #15722
Studio: MOSS-Image GGUF path follows the per-checkpoint negative-branch skip by @danielhanchen in #16255
CI: raise the Core zoo job cap from 35 to 50 minutes by @danielhanchen in #16260
Studio: offloaded diffusion runs both guidance branches in one batched call whenever memory allows by @danielhanchen in #16261
Scan baseline: re-review the twelve unsloth-zoo sites 2026.10.11 moved by @danielhanchen in #16264
Studio: a streamed Qwen-Image load from the hosted int8 checkpoint gets the int8 text stream and modulations by @danielhanchen in #16275
Studio: an offline diffusion load reads the mirror the earlier online load staged by @danielhanchen in #16276
Pin usearch for the document upload harness, and guard its CI-only RAG imports by @danielhanchen in #16284
LoRA: decide the fused order once per call by @danielhanchen in #16282
Studio video: cache the resident runtime-quantized int8 DiT so returning loads skip the requantize by @danielhanchen in #16291
Slot resume tests: fake the disk in every test that calls save_slots_for_resume by @danielhanchen in #16305
Studio: stream the SDXL UNet leaf by leaf when whole-module offload cannot fit it by @danielhanchen in #16295
Studio: build offloaded video int8 denoisers per shard when the dense load does not fit host RAM by @danielhanchen in #16290
RAG ANN test: build the recall check index on one thread by @danielhanchen in #16308
Studio diffusion: keep the streamed text encoder in host RAM on the small-host route when it fits by @danielhanchen in #16304
Studio tests: count the work in the prose-call amortization test instead of timing it by @danielhanchen in #16315
Studio video: run the MiniMax-H3 int8 video decoder on the fused ConvRot act quant and int8 GEMM by @danielhanchen in #16314
Studio: skip the MPS cache clear when torch has no usable Metal device by @danielhanchen in #16326
Studio: the SDXL background compile runs only while no render runs by @danielhanchen in #16312
Studio tests: count an exiting Clef worker as dead while init reaps it by @danielhanchen in #16328
Studio video: keep the share of MiniMax-H3's streamed conditioner that fits in host RAM by @danielhanchen in #16325
Studio: let Transcribe take larger and longer audio on audio.cpp models by @Etherll in #15952
Tests: flush each DDP rank's result before teardown in the plain Trainer test by @danielhanchen in #16357
Studio tests: a frozen background compile may be thawed more than once for one waiting compile by @danielhanchen in #16364
Studio: warm the deferred SDXL speed profile off the render thread by @danielhanchen in #16360
Studio tests: skip the real-prepare sandbox tests on a host without bubblewrap by @danielhanchen in #16371
Notebooks CI: refresh the Colab oracle snapshots to the current GPU runtime by @danielhanchen in #16373
Kaggle CI: install six beside the Kaggle client, and name an import failure as one by @danielhanchen in #16374
Studio: run cheap dataset formatting maps in process for in-memory datasets by @danielhanchen in #16179
Studio: share the loopback SSL context in one-shot llama-server calls by @danielhanchen in #16205
Studio: skip unchanged messages when syncing a chat thread by @danielhanchen in #16213
Studio: walk record offsets and batch member writes in log export by @danielhanchen in #16358
Studio: list saved voices with one folder lookup and one sidecar read each by @danielhanchen in #16354
Studio RAG: compile the docx run XPath once and remember style list lookups per document by @danielhanchen in #16343
Studio: decode native attachment base64 with Uint8Array.fromBase64 when available by @danielhanchen in #16319
Studio: skip pymupdf4llm's table search on PDF pages that cannot hold a lined table by @danielhanchen in #16280
Studio: skip log redaction rules that cannot match a line by @danielhanchen in #16356
Studio RAG: read each spreadsheet number format once per workbook by @danielhanchen in #16339
Studio: stream durable chat events without re-parsing stored payloads by @danielhanchen in #16159
Studio: reuse image header dimensions in the dataset labeling listing while the file is unchanged by @danielhanchen in #16366
Studio: show right-to-left chat messages right to left by @shimmyshimmer in #16348
Studio: split seed Hub file names by string and compile split label matchers once by @danielhanchen in #16352
Studio: keep unchanged Recipe Studio nodes and edges across a drag step by @danielhanchen in #16307
RL: read package versions once per trainer patch pass by @danielhanchen in #16203
Studio: overlap web image lookup with text search and independent URL reads in one tool batch by @danielhanchen in #16174
Studio: list the dataset folder once per upload for the caption-stem check by @danielhanchen in #16353
Studio: update managed Node (node: 24.22.1 -> 24.22.2) by @danielhanchen in #16378
Studio Alpaca formatting: build the missing-input fallback once per batch by @danielhanchen in #16341
Studio: run the docx text rewrites over one unpacked part map instead of rezipping the archive after each by @danielhanchen in #16329
MoE: keep the fitted dW tile for float32 activations in the heuristic configs by @danielhanchen in #15928
Studio Images: encode gallery PNGs about 2x faster with byte-identical output by @danielhanchen in #16190
Studio: ask ismount only where a directory's own stat can show a mount in the sandbox walk by @danielhanchen in #16370
Import: scan installed distributions once while transformers and TRL import by @danielhanchen in #16189
SFT: answer the pre-tokenized cap check from Arrow lengths when they prove it by @danielhanchen in #16293
Studio: read a run's checkpoint state in the history listing only when the answer depends on it by @danielhanchen in #16323
Studio export: upload a reused folder's fresh files instead of saving the model twice by @danielhanchen in #16266
Studio: build the chat search index from a trimmed message view by @danielhanchen in #16207
Studio: resume the Gemma body brace scan from its last end instead of rewalking the open body every token by @danielhanchen in #16346
Studio: stream the recipe batch merge instead of loading a dataframe by @danielhanchen in #16262
FP8: float32 dequant before torch 2.14, no direct path under mismatched autocast by @danielhanchen in #15927
Lean varlen: ask torch 2.15's backend selector instead of the removed _can_use_cudnn by @danielhanchen in #16288
Studio: copy tool schemas without deepcopy in the tool loop controller by @danielhanchen in #16306
Studio: background fit and position controls, outline strength slider, composer setting in Appearance, audio and library polish by @shimmyshimmer in #16287
Studio: cache checkpoint losses per unchanged trainer_state and scan checkpoints off the event loop by @danielhanchen in #16320
Studio diffusion: change LoRA weights without reloading the adapter files by @danielhanchen in #16310
Studio RAG: count chunking tokens in batches by @danielhanchen in #16265
Studio model picker: On Device rows re-render only when their own data changes by @danielhanchen in #16316
Studio: cut repeated work in the local model listing by @danielhanchen in #16146
Studio video: decode reference audio once per request and stage H3 reference frames faster by @danielhanchen in #16324
Studio: stop re-stripping the whole turn per token once tool markup is open by @danielhanchen in #16118
Studio export: save MLX merged and base pushes once into a reused folder by @danielhanchen in #16278
Studio: decode external provider images off the event loop by @danielhanchen in #16362
Studio: reuse the BOS probe encoding instead of tokenizing a long prompt twice by @danielhanchen in #16289
Studio: list the audio gallery from one folder scan and reuse unchanged sidecars by @danielhanchen in #16367
Studio: run diffusion dataset upload file work on a worker thread by @danielhanchen in #16361
Studio: choose where the background image shows (new chat page, chats, or all pages) by @shimmyshimmer in #16297
Studio: greeting toggle in Appearance, and the composer stays put when the greeting is off by @shimmyshimmer in #16301
Studio: Appearance > Code with a live preview, line numbers, word wrap and ligatures by @shimmyshimmer in #16299
Studio: reuse each repo's partial verdict in the local model listing by @danielhanchen in #16199
Studio CI: run the web tools overlap timings serially by @danielhanchen in #16383
Tests: stub the new num_proc helpers in the no-torch venv, and read with utf-8 in the metadata cache test by @danielhanchen in #16386
Studio: reconnect a chat's MCP session after the server restarts by @NilayYadav in #16209
fix(studio): report why a model's security checks blocked it by @Lyxot in #16292
Give Claude its shell back when the local agent can't run a command by @NilayYadav in #16211
Studio: read each gallery PNG's recipe once while it is unchanged and resolve only symlinked clips by @danielhanchen in #16355
Studio: Images and Video strips stop re-rendering every card per thumbnail and page by @danielhanchen in #16311
Studio: long chats stop re-rendering source badges per token and tool cards per run by @danielhanchen in #16277
Studio: keep transcript listing rows while their files are unchanged by @danielhanchen in #16363
Studio: keep <function=...> tool-call parsing linear on long output by @danielhanchen in #16105
Studio Images: another account's loaded model no longer offers Generate by @debojyoti-dev in #16200
Docker: peel annotated tags to their commit when freezing refs by @NilayYadav in #16206
Studio: stop the false "Could not sync pinned chats" toast during a desktop update by @oobabooga in #16188
Studio: stream durable chat run snapshots without re-parsing the stored request by @danielhanchen in #16302
Studio: tell the agent a project's working folder is where its tools run, so it reads the files instead of asking for them by @oobabooga in #16251
Block CUDA graphs: auto mode stays off when decoder blocks span several devices by @danielhanchen in #15931
Studio: cheaper per-token work in web page Markdown extraction by @danielhanchen in #16286
Studio: stop Terminal calls hanging when a command starts a background process by @NilayYadav in #16210
Studio: gate MiniMax-H3 Auto flash on a real flash SDPA kernel (#16294) by @indrajeetapache in #16296
Studio: harden VAE encode dtype for Qwen-Image-2.1 GGUF Edit/Reference (#16375) by @indrajeetapache in #16376
Studio Images: refuse a load whose resolved gated base the account cannot access by @danielhanchen in #16379
Studio: list the video gallery from one folder scan and reuse unchanged sidecars by @danielhanchen in #16372
Docker: stop shipping SSH host keys in the Studio image by @NilayYadav in #16208
Kernel cache key tag and readback-barrier key for every thread (checkpoint recompute, torch >= 2.12) by @danielhanchen in #15997
Flex attention: small shared-memory tiles for short no-grad calls so one short prompt does not leave the process on eager attention by @danielhanchen in #16049
Reset the inference stance counter in for_inference by @danielhanchen in #16197
Studio: say chats and projects are permanently deleted in delete dialogs by @shimmyshimmer in #16390
Studio: train on CPU when no GPU is found, and offer Install Training on GPU-less Chat only installs by @oobabooga in #16175
fix(studio): align settings sidebar rows with search field by @wasimysaid in #16391
Studio MLX: decode reply snapshots incrementally instead of re-decoding every id by @danielhanchen in #16227
Studio: skip the startup manifest migration scan when nothing changed since a clean pass by @danielhanchen in #16270
Studio: keep a backticked tool-call tag in the answer instead of cutting it by @danielhanchen in #16132
Desktop: stop the NVIDIA X11 flicker on WebKitGTK 2.54 (#16318) by @indrajeetapache in #16322
Studio: Glass theme for the sidebar, composer and menus, glass without a background image, and drop the Classic and Minimal themes by @shimmyshimmer in #16333
fix(providers): preserve all Anthropic system instructions by @hsusul in #16369
Studio: stop routing parakeet.cpp GGUFs as chat with a false llama.cpp update hint (#16273) by @indrajeetapache in #16281
Studio: let the background stay visible in open conversations by @danielhanchen in #16381
CLI: list-checkpoints reports an --outputs-dir outside the outputs folder as an error by @hsusul in #16202
Studio: convert ShareGPT image lists in placeholder order by @danielhanchen in #16385
Studio: decode video attachments instead of fetching a data URL by @YakshithK in #16254
Studio: never launch wxc-exec on Windows older than build 26100 by @danielhanchen in #16384
CLI: inference and chat take images for vision models by @debojyoti-dev in #16347
Use tar.xz for Node downloads on macOS and Linux by @NilayYadav in #16253
Studio: serve the Decision API's pinned GGUFs from their ModelScope commits by @debojyoti-dev in #16285
Flex attention on torch < 2.14: fitting tiles for short decode calls so a batch-size change does not leave the process on eager attention by @danielhanchen in #16073
Gemma 3: generate on the non-flex attention path instead of flex_attention's fallback by @danielhanchen in #16041
Decode CUDA graphs: opt-in whole-step capture for batch 1 (UNSLOTH_DECODE_CUDA_GRAPHS_FULL=1) by @danielhanchen in #15976
Studio tests: pin the delete chat placeholder, not its wording by @danielhanchen in #16397
Studio: macOS system glass on Unsloth Desktop by @shimmyshimmer in #16334
Studio: turn on spell checking in Linux Desktop by @danielhanchen in #16380
Studio: keep Trick or Treat as a year-round theme, remove the Halloween extras by @shimmyshimmer in #16392
Studio: run and hide MiniMax-M2 tool calls that llama-server leaves in content by @danielhanchen in #16103
Studio: answer a collapsed On Device row from disk, not the Hub by @yannickloth in #16271
Studio: exit the inference worker instead of waiting to be killed by @YakshithK in #16338
CLI: let train load models that need trust_remote_code by @YakshithK in #16332
Studio: respawn a llama-server that ran out of GPU memory mid-session with more VRAM headroom by @danielhanchen in #16388
Studio tests: data settings driver expects the named single-chat delete for one archived result by @danielhanchen in #16401
Unsloth Studio MLX: Fast Verify toggle by @Lyxot in #16259
FSDP2 float16 with CPU offload: unscale CPU DTensor grads as local shards by @danielhanchen in #16398
Studio: decode with CUDA graphs by default in the transformers chat backend by @danielhanchen in #15993
Studio tests: compare the hub source restore against the inherited source by @danielhanchen in #16236
Stub the device context in the packed fp32 routing test by @danielhanchen in #16229
GRPO PrefixGrouper: check the host-syncs layout test against the reference by @danielhanchen in #16244
Tests: GQA routing oracle expects expanded K/V on torch without enable_gqa by @danielhanchen in #16225
Studio: search earlier turns for the user's question, not the tool loop's own notices by @debojyoti-dev in #16204
Studio: make Show in folder work on Linux, and report when it can't by @debojyoti-dev in #16198
Tests: pin bf16 support and the HF cache location in the GGUF preflight fixtures by @danielhanchen in #15967
Add unsloth.precision: one registry for low-precision recipes, refusals and high-precision layers by @danielhanchen in #16387
Tests: run the RoPE 64-bit indexing metadata check on CPU-only runners by @danielhanchen in #16226
Studio: use the macOS system font (SF Pro) as the default UI font on Mac by @shimmyshimmer in #16402
Studio: customizable new-chat greeting by @shimmyshimmer in #16408
Studio tests: skip the CUDA compiled-linear checks where production stays eager by @danielhanchen in #16237
Studio: use the Hugeicons trash and bookmark-01 icons by @shimmyshimmer in #16409
Studio tests: a separate engine port block for every xdist worker past 40 by @danielhanchen in #16245
Full finetuning notice: say float32 for forced-float32 models on GPUs without bfloat16 by @danielhanchen in #16405
Studio: run multi-GPU compile tests on real CUDA, fail on eager fallback by @danielhanchen in #16224
Block compile: install the router_logits capture hooks before the first block compiles by @danielhanchen in #16410
Pipeline parallel: carry the Qwen3-MoE router auxiliary loss across stages by @danielhanchen in #16406
Tests: gpt-oss dense-expert reload fixture drops replaced experts' quant state by @danielhanchen in #16246
Studio sandbox: remap the Windows spelling of /mnt/data-style paths by @danielhanchen in #16412
Studio tests: best of five for the two single-sample nested-header linearity checks by @danielhanchen in #16426
Risky loader baseline: record the float8 training precision fallback by @danielhanchen in #16428
Studio tests: take the hub source baseline when the first test starts, not at collection by @danielhanchen in #16430
Studio tests: chat width driver picks the presets by position, not by their copy by @danielhanchen in #16431
Studio: shorten the Hugeicons trash and bookmark comments by @shimmyshimmer in #16439
Multi-account Playwright: record and retry once when the post-logout /login load never fires by @danielhanchen in #16437
Profile stats tests: seed at noon, so a run just before midnight keeps the reply on today by @danielhanchen in #16440
Studio: update Hugeicons to 4.4.0 by @shimmyshimmer in #16442
UNSLOTH_PARALLEL: one flag and one root mesh for dp / tp / pp / cp, pipeline parallel from the environment, and a capability registry by @danielhanchen in #16407
Studio: replace stray Lucide icons and settle duplicate icons by @shimmyshimmer in #16445
Repair the three CI checks #16407 left red on main by @danielhanchen in #16446
Studio: read lookup tables lazily when needed to enable the MoE cache by @oobabooga in #16444
Studio: turn navigator at the chat's left edge, building on #11668 by @shimmyshimmer in #16300
PP env GA test: compare the trained adapter at 5e-5 by @danielhanchen in #16457
Studio: add a Time format setting and drop the redundant desktop update note by @shimmyshimmer in #16448
Backend CI: run the context parallel compose gloo suite in its own shard by @danielhanchen in #16447
Preserve external chat import recovery across checkpoint failures by @danielhanchen in #16458
Studio: list every Gemini chat model (follow nextPageToken, allow new families) by @debojyoti-dev in #16433
Studio: estimate the prompt's time left from how the cost per token grows, not the rate so far by @oobabooga in #16450
Studio: keep the live generation line from moving the branch picker, and update it during short replies by @oobabooga in #16449
Studio: time a continued reply as the whole answer, not just its last pass by @oobabooga in #16451
Studio: Composer appearance section, composer outline on by default by @shimmyshimmer in #16455
Studio: translate the live generation figures, the reply timing tooltip and the weight placement row by @oobabooga in #16452
Studio: add a link preview setting by @danielhanchen in #16459
Fix OpenRouter MAI-Voice speech playback by @danielhanchen in #16463
fix(api): preserve refusal text in collected connection replies by @hsusul in #16435
Studio: honour Voice Settings Load into on cold STT API loads (#15280) by @indrajeetapache in #15284
studiobench parity: treat every data-*-id attribute as an instance id by @danielhanchen in #16464
Studio: report the measured VRAM shortfall instead of blaming the previous worker by @YakshithK in #16453
Fix pipeline denoiser downloads labeled as required assets by @Imagineer99 in #16465
Recognize shared encoder assets in Model Hub inventory by @Imagineer99 in #16466
Studio: paste links, tables, lists and formatting into the composer as Markdown by @shimmyshimmer in #16298
Studio: keep the chat background, interface scale and Browser preferences with the account by @shimmyshimmer in #16454
Studio: keep a text selection in Settings › Logs across polls so part of the log can be copied by @LeoBorcherding in #16099
Studio: match diffusion training access facts to the no-token mirror fetch by @danielhanchen in #16216
Studio: /v1/chat/completions puts the whole answer in reasoning_content for templates with a {% generation %} tag by @LeoBorcherding in #16121
Skip pinned staging for native int8 host quantisation on integrated GPUs by @danielhanchen in #16235
Studio: restore public OpenAPI descriptions on four routes and schemas by @danielhanchen in #16223
Studio: apply the mapped chat template once per loaded tokenizer by @danielhanchen in #16377
Studio: train audio VLMs on the images in a mixed audio + image dataset by @LeoBorcherding in #16124
Studio: keep feature indexes and the markdown renderer out of the startup bundle by @danielhanchen in #16394
Studio: keep a nested active HF hub clearable and listed by @danielhanchen in #16215
Concretize RMSNorm width before the narrow-row self-check by @danielhanchen in #16228
Fused LM head: accept TRL 1.17's score_ids kernel argument by @danielhanchen in #16220
Block compile: say when packed rows run each block uncompiled by @danielhanchen in #16272
Studio video: reuse bounded conditioning image decodes by @danielhanchen in #16403
Studio: show the voice description on Clone for Breeze TTS 2 and VoxCPM2 by @Etherll in #15878
Studio: do not subtract live segment graph outputs twice from reusable cache by @danielhanchen in #16231
Studio: keep an anonymous Hub token anonymous in the hosted bf16 VAE path by @danielhanchen in #16214
Studio: wait on worker exit instead of fixed sleeps when switching models by @danielhanchen in #16344
Studio Windows XPU Triton: verify the swap and torch's XPU triton payload, repair a broken triton, skip pip backups by @danielhanchen in #16250
Studio: wake the dispatcher on stop instead of waiting out its queue poll by @danielhanchen in #16350
Studio: recognise a worker VRAM release that finished before the first settle poll by @danielhanchen in #16404
Saving: drop the training truncation and padding from saved tokenizers by @LeoBorcherding in #16102
Cleanup: remove six private helpers and constants in core that nothing calls by @danielhanchen in #15695
Desktop: overlap owner liveness probes and skip redundant preflight HTTP by @danielhanchen in #15739
Studio: name FFmpeg when a video GGUF can't decode, and shrink clips with the ffmpeg llama-server uses by @LeoBorcherding in #16098
Studio: build the code highlighter from shiki/core so the unused WASM engine leaves the bundle by @danielhanchen in #15740
Studio diffusion: turn off coordinate descent where Max pins reduction configs by @danielhanchen in #16269
Studio: start the credential bootstrap with the entry instead of after the route checks by @danielhanchen in #16389
Loader: fetch each config probe once per from_pretrained call by @danielhanchen in #15757
Saving: write the source tokenizer_class instead of TokenizersBackend by @LeoBorcherding in #16104
Studio: LFM3-Omni audio input, export, LoRA defaults and Stop for custom-code generate by @LeoBorcherding in #16242
feat(studio): draft MLX tool, video, grammar, KV and LoRA replies by @Lyxot in #16441
Studio sd.cpp install: treat ELF symbol-version mismatches as loader refusals by @danielhanchen in #16150
Studio fine-tuning queue: do not start jobs cancelled during the batching probe by @danielhanchen in #16151
Studio: keep style keywords inside quoted state labels in the Mermaid guard by @danielhanchen in #16133
Studio: emit only the woff2 KaTeX fonts by @danielhanchen in #16382
FP8: strip activation_scale sources when matching the 16-bit save conversion by @danielhanchen in #16156
Studio: chat with LFM3-Audio-2B (text and speech in, text out) by @LeoBorcherding in #16243
Decode graphs: preserve eager cache growth and reject stale replay by @danielhanchen in #16191
Studio OpenAI API: refuse unsupported weights in selected JSONL chats by @danielhanchen in #16153
Flex small-smem tiles: keep float32 calls on the default tiles by @danielhanchen in #16172
Prevent Desktop owner lockout before adding accounts by @mahiatlinux in #13961
Studio: carry FP8 / NVFP4 compile cache tags into render and compile threads by @danielhanchen in #16158
chore: pre-commit autoupdate by @pre-commit-ci[bot] in #16492
FSDP2: average tp gradients over dp when every trainable param is ignored by @danielhanchen in #16180
Studio video: load cached weights offline while the Hub is unreachable by @danielhanchen in #16152
Packing: do not serve compiled inference_mode metadata to a training step by @danielhanchen in #16183
Studio diffusion prequant cache: do not count cgroup shmem as reclaimable by @danielhanchen in #16154
FSDP2 save: one writer for a tensor parallel x FSDP2 model by @danielhanchen in #16181
FSDP2 QAT export: count the converted copy in the writer's host RAM check by @danielhanchen in #16166
Desktop: recheck the install fingerprint before accepting a cached capability answer by @danielhanchen in #16134
Studio Compare: keep the negative prompt on cells the Detailer pass still uses by @danielhanchen in #16149
Studio diffusion: a numbered shard without its index is not a complete snapshot by @danielhanchen in #16418
Multi-LoRA grouped experts: run the grouped support and parity probes only eagerly by @danielhanchen in #16415
Packed mRoPE attention: no graph break on warm packed rows by @danielhanchen in #16416
Resolve compatible Unsloth and Zoo pairs for all no-torch installs by @danielhanchen in #16157
Fix FSDP2 checkpoint planning across parameter swaps and accumulation by @danielhanchen in #16171
SDPA training on SM100: keep the caller's backend policy when avoiding cuDNN by @danielhanchen in #16230
Studio Hub download: keep refs/main on the last complete snapshot under failed updates and concurrent workers by @danielhanchen in #16422
Keep fused-loss engagement counters out of traced code by @danielhanchen in #15575
PPO fused rollouts: keep TRL's scores when generate applies logits processors by @danielhanchen in #15574
Studio: keep a generated chat title in the conversation's language by @debojyoti-dev in #16413
Studio: save idle auto-unload seconds on Enter by @NilayYadav in #16501
Studio: accept auto and all in custom GPU-layer arguments by @danielhanchen in #16518
Fix emoji reading in chat read-aloud by @rugved8607 in #16480
Studio: darken Liquid Glass 3D composers over bright wallpapers by @wasimysaid in #16477
Studio: offer only the reasoning levels a template allows and show its raise message by @oobabooga in #16502
Recover VAE storage after precision-conversion OOMs by @danielhanchen in #16147
Studio settings: size the composer Beta badge and the settings search offsets from the UI scale by @danielhanchen in #16524
Worker stderr mirror: read once more after an empty read that raced the stop by @danielhanchen in #16525
Parallel isolation guard: list the Hub download tests poll deadlines as benign timing by @danielhanchen in #16526
Save Desktop Gallery batches with one folder choice by @danielhanchen in #16521
No-Triton import: hand the triton stub to unsloth importers without publishing it by @danielhanchen in #15576
FSDP2: sum fp16 replicated gradients over unsynced micro-batches in float32 by @danielhanchen in #16173
Scoped download tests: seed the Hub cache the offline path now pins by @danielhanchen in #16530
Multi-LoRA cold compile test: read the source as utf-8 by @danielhanchen in #16531
No-torch install harness: model a feed that resolves the Unsloth / Zoo pair by @danielhanchen in #16532
Pipeline parallel: end a non-last stage without a graph break under torch.compile by @danielhanchen in #16167
No-deps pair test: start PowerShell through the shared runner by @danielhanchen in #16533
FP8: keep rowwise _scaled_mm for float32 where cuBLAS supports it (torch < 2.14) by @danielhanchen in #16393
GPU supervisor signal test: publish its handshake files atomically by @danielhanchen in #16535
Pipeline parallel: agree on overflow once when explicitly unscaling by @danielhanchen in #16417
Prevent FP16 overflow in grouped multi-LoRA expert projections by @danielhanchen in #16414
MLX inference stub: answer the _FIRST_USE_WORK import zoo added by @danielhanchen in #16537
ROCm progress test: model a feed that resolves the no-torch pair by @danielhanchen in #16542
Multi-LoRA overflow test: read the kernel source as UTF-8 by @danielhanchen in #16544
Studio: combine OpenCode Console and Go into one connection by @NilayYadav in #16482
Fix Connections dropdown shrinking on scroll and hover by @wasimysaid in #16543
Studio: size audio.cpp TTS models from measured peaks and run TQ2_1 on the CPU by @oobabooga in #16514
Desktop: allow a custom backend port with UNSLOTH_DESKTOP_PORT by @danielhanchen in #16523
Studio: keep the loaded model when Switch back has to download a GGUF by @NilayYadav in #16474
Respect interface scaling in Settings search and rich-paste badge by @danielhanchen in #16471
Block Swap: get_peft_model crashes on MoE models with offload_layers (evicted expert weights read as empty) by @LeoBorcherding in #15159
Keep a verified live Desktop backend after watchdog recovery by @danielhanchen in #16470
Keep Studio windows signed in during concurrent session refresh by @danielhanchen in #16468
Studio diffusion: skip settling for proven resident replans by @danielhanchen in #16400
Make page-in startup progress tests independent of host scheduling by @danielhanchen in #16476
Preserve prompt cache during deferred MCP tool search by @wasimysaid in #16546
Studio: show image generation time by @Aafimalek in #16462
Preserve split-device compressed-tensors input transfers by @danielhanchen in #16155

New Contributors

@yannickloth made their first contribution in #16271
@rugved8607 made their first contribution in #16480

Full Changelog: v0.1.913-beta...v0.1.914-beta

unslothai/unsloth v0.1.914-beta Glass Theme, Personalize & Faster Generation on GitHub