unslothai/unsloth v0.1.912-beta on GitHub

Halloween 👻 is here. Through October, a Trick or treat row in the sidebar opens eleven Halloween themes, spooky greetings and a few spiders and bats.

This release also speeds up the first launch, Linux tool calls and model loading, and trains compatible fine-tuning jobs together with multi-LoRA.

Highlights

Trick or treat: 🎃 new Halloween themes - until Nov 1
Faster Unsloth: 2x faster first launch, plus speedups for tool calls, model lists, dataset prep and MiniMax-H3 loads
Multi-LoRA batching: up to 4 compatible queued fine-tuning jobs train together
SGLang thinking + auto-compact: Resume and JSON output now work on vLLM and SGLang too
Browser use switch: let models control the built-in browser from its new tab page

Optimizations

Setup now precompiles Python files, so the first launch is about 2x faster.
On Linux, tool calls start in 1 to 2 ms instead of 84 to 174 ms with a 2 GB server process.
With 497 cached models, the model list uses about a third of the CPU time it did.
Preparing a 1,000-row chat dataset for Gemma 3 4B took 1.75 s instead of 6.6 s in our test.
MiniMax-H3 loaded in 49 s instead of 61 s on a Linux A100.

Multi-LoRA + API

When several compatible jobs wait in the fine-tuning queue, Unsloth now trains them together, each with its own LoRA adapter on a shared base model.

Jobs sent while the GPU is busy queue up in order, and each can be cancelled on its own.
Load a job's fine_tuned_model name to chat with its trained adapter.
/v1/videos requests now queue while a clip renders, and matching /v1/images/generations requests are batched.

How it works

New Triton kernels run the base layer once per batch, while each adapter learns only from its own tokens with its own loss and learning rate.
With 4 adapters, a 4096x4096 layer's forward and backward pass ran 1.52x faster than one adapter at a time (A100, bf16, rank 16).

Chat + tools

Local models skip tool calls that hit the output limit, so a half-written cat > app.py can't overwrite a file.
Before, cut-off rewrites shrank a 175-line test file to as few as 6 lines on an RTX 3090.
Thinking models such as Qwen3 and DeepSeek R1 now show their reasoning and the Think toggle on SGLang.
With Auto-compact on, long SGLang chats drop their oldest turns and keep answering.
On vLLM and SGLang, you can resume a reply cut off at Max Tokens and request JSON with response_format.

Browser + Desktop

The browser's new tab page has a Browser use switch that lets models control the browser.
On Desktop, a browser agent working for a temporary chat no longer adds pages to your history.
On macOS with Safari 18.4 or earlier, Desktop browser tabs and their close buttons work again.
Log out is now in the Unsloth Desktop user menu.

Trick or treat

The Trick or treat… row sits above More in the sidebar and opens a Halloween menu.

Pick one of eleven themes, such as Pumpkin, Coven and Ghost Ship, each with a light and dark mode.
Turn on Spooky greetings and new chats open with a Halloween line and sloth.
With Spiders and bats on, Halloween themes get the odd spider and bats crossing the window.
Not sure which theme? Trick or treat! picks one at random.

Unsloth returns to normal at midnight on November 1, or right away with Remove now.

Training

Gemma 4 E2B and E4B now keep a dataset's thinking text in training instead of dropping it.
New chat training setups turn on sequence packing and Assistant completions where the model and data support them.
Local JSON and JSONL datasets load even when later rows add new keys.

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

Platform	Link
Windows	Download
macOS	Download
Linux x64 / Ubuntu (deb)	Download
Linux ARM64 / Ubuntu 24.04+ (deb)	Download
Linux x64 (AppImage)	Download
Windows ARM64	Download

What's Changed

Studio Images: no toast when the prompt upsampler is not available by @danielhanchen in #15365
Bump install.sh / install.ps1 pins to unsloth>=2026.10.9 by @danielhanchen in #15380
TP Trainer parity tests: run AdamW with eps 1e-4 so reduction-order noise is not amplified by @danielhanchen in #15381
Studio: price per-layer embedding and Engram lookup tables as host or disk memory, not VRAM by @danielhanchen in #14465
Studio video: refuse NaN / inf decodes for every family before a PIL / uint8 cast or clamp hides them by @danielhanchen in #15374
Prompt queue: row button labels do not take the next row's clicks by @danielhanchen in #15383
Studio diffusion: pre-seed the background compile's float specialisations on a returning load by @danielhanchen in #15393
Studio sd.cpp: verify a new bundle starts before it replaces the installed one; Windows ROCm without the HIP SDK uses Vulkan by @danielhanchen in #15371
TP Trainer parity tests: keep eps 1e-4 to the UNSLOTH_TP_SIZE runs, native Trainer runs back on the default by @danielhanchen in #15397
Multi-account Playwright: retry the login page once when the app's own logout redirect aborts it by @danielhanchen in #15398
Studio video: refuse a NaN render on the native MiniMax-H3 route by @danielhanchen in #15394
TP x FSDP2 parity test: per-arch step-norm tolerance for phi3's fused shards by @danielhanchen in #15400
FSDP2 / DDP: only keep torchao weights FSDP2 cannot shard out of fully_shard and the DDP broadcast by @danielhanchen in #15373
Studio backend tests: close a scoped MCP session a test leaves cached by @danielhanchen in #15407
Keep SFT loss_type="dft" on the DFT objective when label smoothing is set by @danielhanchen in #15396
loader: do not restore an fp8 quantization_config over dequantized weights by @danielhanchen in #15364
FSDP2 bitsandbytes: no per-parameter device sync, bounded GPU memory in the optimizer save by @danielhanchen in #15384
Tensor parallel: generate, adapter / tokenizer saves, merged and GGUF exports, and reload cells under UNSLOTH_TP_SIZE by @danielhanchen in #15368
Tensor parallel: average fp16 dp gradients in fp32 so finite gradients cannot sum to inf by @danielhanchen in #15415
FSDP2: gather before clearing the post-forward mesh in native_fsdp2_unsharded by @danielhanchen in #15426
gpt-oss: read config.transformers_weights before the standard filenames when probing the expert layout by @danielhanchen in #15386
Native FSDP2 fp16 scaler: agree on overflow over ranks without an explicit unscale_, and give fused optimizers one 0 / 1 flag over all devices by @danielhanchen in #15416
Attention: packed fp32 rerouted to xformers keeps an fp32 output without autocast by @danielhanchen in #15410
Tests: fix the two spoof-safety lint failures on main by @danielhanchen in #15454
Studio tests: queue each FIFO admission waiter before starting the next by @danielhanchen in #15457
Studio video: pre-seed the denoiser compile's float specialisations on a returning load by @danielhanchen in #15406
Keep DeepseekV4 per-label rope bases when rope_scaling is replaced on transformers 5.21 by @danielhanchen in #15425
Studio: centre the chat column under classic scroll bars in WebKit by @shimmyshimmer in #15392
Studio: a switch per permission mode in Settings > Sandbox by @shimmyshimmer in #15469
Studio: Run settings divider, preset picker fill and a clearer Always load by @shimmyshimmer in #15472
Studio: top-align the settings rows the shared row components do not cover by @shimmyshimmer in #15468
Studio: Library chat, project and section cards keep one size by @shimmyshimmer in #15471
Studio: solid primary buttons on update notifications by @shimmyshimmer in #15470
Studio: API examples behind one tab row, an access group, API page fixes by @shimmyshimmer in #15467
Studio: Audio Edit step numbers as text, standard Transcribe icon by @shimmyshimmer in #15484
Studio: size the Transcribe icon from its label by @shimmyshimmer in #15486
Studio diffusion: pin Qwen-Image-2.1's reduction-config filter (same seed, same image across processes) by @danielhanchen in #15455
Save: compare fused 4-bit expert bytes bitwise for float quant_storage by @danielhanchen in #15378
Version compat: transformers 5.21.0 is tagged v5.21 upstream by @danielhanchen in #15438
Studio tests: turn Hub telemetry off where mocked Hub requests are counted by @danielhanchen in #15453
Studio tests: probe int8 CPU quantize + flatten apart from the cache assertions by @danielhanchen in #15459
FP8 block decode: keep the fused small-M kernel for fp16 when torchao is selected by @danielhanchen in #15361
QAT: fake quantize embeddings under FSDP1 and torch.func.functional_call by @danielhanchen in #15385
Routing replay: score named PEFT reference adapters with their own routing by @danielhanchen in #15387
fast_inference under torchrun: every rank takes the same vLLM-or-HF path by @danielhanchen in #15395
Settings contract tests: follow MonitorLink into the API page's access group by @danielhanchen in #15500
FSDP2 sharded loading gloo test: barrier before a rank tears down its group by @danielhanchen in #15499
MXFP4 packed matmul: bound the dX decode like the forward below sm_80 by @danielhanchen in #15376
Studio: orange Full access in permission menus, sandbox docs link, Confirm by @shimmyshimmer in #15504
Studio video: size the VAE decode from free memory after the denoiser offload (wider tiles, same clip whether the CUDA graph is kept or dropped) by @danielhanchen in #15464
Studio: keep timestamps off for every GPT transcription model on a connection by @danielhanchen in #15437
for_training flag test: lift every _utils helper for_training reaches by @danielhanchen in #15493
Studio video: size the hosted H3 text encoder before the unified-memory guard by @danielhanchen in #15474
Studio engines: certifi 2026.10.19 in the vLLM and SGLang locks by @danielhanchen in #15495
tests: stub the tensor parallel guard in the for_training extraction harness by @danielhanchen in #15498
Studio: use shared download manager for diffusion training by @danielhanchen in #15476
Precision conflict test: lift every loader.py helper from_pretrained reaches by @danielhanchen in #15515
MXFP4 low memory test: gate the chunked dX budget test on a real CUDA device by @danielhanchen in #15516
Add batched multi-LoRA training kernels by @danielhanchen in #14648
Studio: classify host pinning by the target device, not the current one by @danielhanchen in #15463
Studio: account hosted quantized components of every image and video family in cache cleanup by @danielhanchen in #15487
Studio ROCm: tell self from cross attention without guarding the query length on the key length by @danielhanchen in #15458
Scan baseline: re-review the seven unsloth-zoo sites 2026.10.9 moved by @danielhanchen in #15524
fix(studio): center sidebar header icons in hover circles by @wasimysaid in #15528
Studio: tighter gap under the Chat with files header by @shimmyshimmer in #15523
Keep Mamba x_proj and out_proj unquantized for the fused mamba_ssm training path by @danielhanchen in #15377
Studio diffusion: keep the compiled int8 W8A8 linear at or below the eager peak, honor force-eager by @danielhanchen in #15456
Studio: Video Create and Train tabs, media header icons, Images and Video linked by @shimmyshimmer in #15520
Studio: More menu hover pill and row height, grouped chat options menu by @shimmyshimmer in #15519
Studio: media rails keep their width when the scrollbar appears by @shimmyshimmer in #15522
Studio: preset and Inpaint icons, aligned preset save row, component file fields at select height by @shimmyshimmer in #15530
fix(studio): center header menu dots without enlarging them by @wasimysaid in #15534
Full finetuning: dequantize block FP8 checkpoints on load by @danielhanchen in #15367
FP8 TMA block GEMM: cut per-call descriptor host cost by @danielhanchen in #15372
float8_training: keep Accelerate dispatch hooks on converted layers by @danielhanchen in #15363
CPU backend: export load_ebook_dataset, RawTextDataLoader and TextPreprocessor by @danielhanchen in #15501
Flex attention: keep float32 Gemma 3 off the sm86 / sm89 small tiles by @danielhanchen in #15401
Packed attention: recheck the SDPA policy before serving cached varlen metadata by @danielhanchen in #15402
Studio: Browser use switch on the browser's new tab page by @shimmyshimmer in #15540
Studio: heading for the API monitor card in Settings > API by @shimmyshimmer in #15544
Fused LoRA: an unfrozen plain Linear after merge_and_unload / unload runs its own forward by @danielhanchen in #15399
RoPE override: normalise the legacy type key on per-layer-type configs by @danielhanchen in #15403
FSDP2: keep the use_cpu reclaim reachable when only import unsloth wrapped the Trainer by @danielhanchen in #15427
Studio diffusion: do not cache a VAE rescaled for fp16 decode by @danielhanchen in #15511
Studio: recognise model.safetensors-N-of-M shards for training resume (#15405) by @indrajeetapache in #15408
Risky loader baseline: record the CPU backend dataprep loader from #15501 by @danielhanchen in #15549
Keep a wrapped Inductor pass's None uuid as a cache opt-out by @danielhanchen in #15429
FSDP2 SentenceTransformer save: honour create_model_card=False and keep training mode by @danielhanchen in #15431
Keep Unsloth Studio data when following the toolkit installer's command by @NilayYadav in #15450
Studio diffusion: keep SDXL ControlNet running after the T4 resident promotion by @danielhanchen in #15446
Desktop frontend contract: follow the media header links to their icon buttons by @danielhanchen in #15550
NVFP4 QAT: keep the fake-quantized linear out of compiled graphs by @danielhanchen in #14812
Packed attention: run the in-graph varlen probe with grad enabled, so a checkpoint's no_grad first trace keeps the varlen path by @danielhanchen in #15414
Notebook sync durability test: pick the abandoned-restore victims in the state's own order by @danielhanchen in #15548
Studio diffusion: shipped activation seed for Qwen-Image-Edit placement by @danielhanchen in #15542
Studio: neutralize DeepSeek-V4.1's image marker in replayed assistant text by @debojyoti-dev in #15288
Studio: revealing a chat's sandbox folder is owner-only by @danielhanchen in #15512
Studio: size MIG CUDA instances from their own slice, never a physical card row by @danielhanchen in #15465
Studio video: drain palettegen per frame in the two-pass GIF export by @danielhanchen in #14958
Studio training: describe public mirror access separately from model licensing by @danielhanchen in #15475
Studio video: MiniMax-H3 cache-only loads name the missing file and how to fix it by @danielhanchen in #15547
QK norm + RoPE backward: pin the dX contraction so compiled fp16 matches eager on ROCm by @danielhanchen in #15506
Block compile: mark Unsloth's own compiled region apart from a user's enclosing compile by @danielhanchen in #15526
Studio: keep browser tabs clickable on macOS with Safari 18.4 and earlier by @danielhanchen in #14368
Refuse Trainer(model_init=...) under UNSLOTH_TP_SIZE instead of silently skipping tensor parallel by @danielhanchen in #15424
Studio diffusion: fused RMSNorm on the fp16 eager Z-Image path by @danielhanchen in #15543
Studio diffusion: keep reference headroom on multi-GPU FLUX.2 klein splits so edits are not refused by @danielhanchen in #14858
cuDNN SDPA backward: give grad_output the output's layout on torch 2.6 - 2.13 by @danielhanchen in #15445
Studio: show the chat Seed setting for vLLM models by @NilayYadav in #15432
Studio: refuse a reused temp folder that other users can still read by @danielhanchen in #14962
Studio native cache write: read each weight from the live module at write time so offload can reclaim VRAM by @danielhanchen in #15461
Studio: queue OpenAI fine-tuning jobs while the GPU is busy by @danielhanchen in #14645
Studio: thinking support for models served by the SGLang engine by @NilayYadav in #15433
Studio diffusion: store chunked int8 W8A8 results straight into the output rows by @danielhanchen in #15533
Studio: keep a temporary chat's browser agent pages out of history by @danielhanchen in #14956
REAP: accept LoRA models merged on PEFT 0.18 by @danielhanchen in #14850
Make the one scored completion GRPO test follow the TRL version by @danielhanchen in #14775
MXFP4 dequantize branch test: run the locals the branch reads ahead of it by @danielhanchen in #15558
Studio: respell user ngram-mod extra args for older llama.cpp builds by @danielhanchen in #14880
fix(studio): time a speculative vision batch row's prefill from before add() by @Lyxot in #15517
Studio: start overlapping document chunks on a whole word by @NilayYadav in #15442
Studio: update managed Node to 24.22.1, apply new pins on every update path, and keep the pin current by @oobabooga in #15423
Keep each unsloth eval run's results instead of overwriting the last by @NilayYadav in #15451
Studio: track MPS denoise progress with torch.mps events by @Lyxot in #15323
fix(studio): price DeepSeek-V4.1 oQ quants as unsloth-zoo loads them by @Lyxot in #15508
Studio engines: verify TLS against Studio's certifi bundle by @danielhanchen in #15509
Studio: keep llama-bench in the llama.cpp install by @danielhanchen in #14881
Studio: keep chat-installed packages out of plugin discovery on approved host calls by @danielhanchen in #14948
Studio: keep sidebar sections and favorites made in another window by @NilayYadav in #15440
feat(studio): store ComfyUI fp8 layers as fp8 on Apple MPS when the widened model would not fit by @Lyxot in #15507
Studio: enable packing and assistant-only training defaults where supported by @danielhanchen in #15478
Studio mermaid guard: a quoted @{ cannot hide a later remote image by @danielhanchen in #14907
Studio: keep documents searchable after upgrading a Chat only install by @NilayYadav in #15443
Studio video: map MiniMax-H3's hosted int8 checkpoint instead of reading it whole (faster load, same weights) by @danielhanchen in #15563
Studio: rewind files more than four folders deep by @NilayYadav in #15439
Studio video: pin MiniMax-H3's reduction-config filter (same seed, same clip across processes) by @danielhanchen in #15564
Studio: train compatible queued fine-tuning jobs together with multi-LoRA by @danielhanchen in #15344
feat(studio): show MLX versions in Settings by @Lyxot in #15369
GRPO kept-set refusals: drop the automatic-default wording by @danielhanchen in #15531
Studio: keep the MLX chat context when an image or video model takes the GPU by @Lyxot in #15358
Studio: keep linked folders syncing when files become ignored by @NilayYadav in #15441
Studio: resume cut-off replies and structured output on vLLM and SGLang by @NilayYadav in #15435
Keep Codex MCP servers when running unsloth start codex by @NilayYadav in #15452
Unsloth: show complete chat error messages by @Imagineer99 in #15496
Studio diffusion: pin the up-front slabs in registered host memory (faster load, same bytes) by @danielhanchen in #15562
cuDNN SDPA DTensor test: gate on a real CUDA device by @danielhanchen in #15582
fix(studio): make room for the MLX drafter in the auto offload plan by @Lyxot in #15342
Studio: chat with a fine-tuned model by the name its job returns by @NilayYadav in #15449
Studio: auto-compact long chats on the SGLang engine by @NilayYadav in #15434
Studio video: price deferred pins at their bytes so a returning Wan load pins its text encoder by @danielhanchen in #15612
gpt-oss: refuse fused 4-bit with dtype float32 on GPUs without bfloat16 by @danielhanchen in #15614
Studio: leave room for the whole reply cap in the estimated managed-engine fit by @danielhanchen in #15610
Persist chat pins across clients of the same Unsloth server by @Imagineer99 in #15293
Video encoder pin test: read video.py as UTF-8 by @danielhanchen in #15623
Studio: ask before Clear all deletes the desktop transcript history by @NilayYadav in #15570
Tests: stop pytest 9.2 monkeypatch from deleting torch config entries, and name any test that still does by @danielhanchen in #15626
Import hoist verifier: a local import re-pointed after its definition moved is not a clash by @danielhanchen in #15622
fix(studio): reject standalone GGUF adapters before model teardown by @Aafimalek in #15421
Backend CI: pin pytest below 9.2 by @danielhanchen in #15624
Studio: Qwen-Image-Edit streamed-tier activation seed keeps more of the DiT resident on tight cards by @danielhanchen in #15645
Studio: Library cards stay apart at every zoom, one size, 5 per row by @shimmyshimmer in #15643
Studio: audio field titles stop focusing their field, quieter Optional tags by @shimmyshimmer in #15651
Tests: fail grouped tests by name when --dist loadgroup stops grouping (pytest 9.2) by @danielhanchen in #15638
Studio: Export page in the Train style, Train polish, hide Q4NX without an AMD NPU by @shimmyshimmer in #15640
Studio: rounded browser tabs and hovers, Recents above Suggested once visited by @shimmyshimmer in #15649
Keep LoRA adapters on Mamba x_proj / dt_proj / out_proj when the fused training kernel is available by @danielhanchen in #15539
Run each FlashAdamW update on its parameter's CUDA device by @danielhanchen in #14773
Fix the FSDP2 QLoRA non-float parameter guard on Python 3.9 by @danielhanchen in #14774
Tensor parallel float16: unscale DTensor and plain grads apart in the native ShardedGradScaler by @danielhanchen in #15551
Block CUDA graphs: run MoE layers and flex attention blocks compiled without graphs by @danielhanchen in #15553
Flash impl: warn when an explicit fa3 / fa4 request falls back on the device or torch by @danielhanchen in #15554
Context parallel and async GRPO: recognise gated fast forwards by @danielhanchen in #15559
Scan baseline: triton 3.9 compile warmup pool by @danielhanchen in #15658
checkpoint_skip_layers="auto" under torchrun: every rank keeps the smallest planned count by @danielhanchen in #15545
Studio: recheck that the user is still away before an automatic task swaps models by @danielhanchen in #15436
Studio: expanded composer fills the chat area, net of attachments by @shimmyshimmer in #15650
Studio: restore the reload icons replaced in #15232 by @shimmyshimmer in #15648
Studio: Scheduled task dialog padding, pill name field, dark New task button by @shimmyshimmer in #15647
Fast decode: follow a model moved after loading (GRPO on a non-default GPU raised cuda:1 and cuda:0) by @danielhanchen in #15677
Distributed tests: take every MASTER_PORT from the shared rendezvous helper by @danielhanchen in #15679
Studio: count the cards a CUDA + Vulkan llama.cpp reaches only through Vulkan by @debojyoti-dev in #15466
Condense module docstrings, keeping the guarantees they state by @danielhanchen in #15594
install-kernels: install mamba_ssm on sm70 with Triton 3.4+, matching unsloth_zoo by @danielhanchen in #15557
Keep a dataset's thinking text when training Gemma 4 E2B and E4B by @NilayYadav in #15569
Studio: say a chat is too long on safetensors models by @NilayYadav in #15616
Studio: refuse Laya questions whose instructions would be cut by @NilayYadav in #15602
Studio desktop: never let the login-shell PATH probe hold back launch by @danielhanchen in #15567
Studio: load mermaid, node-forge and js-yaml on first use, not at startup by @danielhanchen in #15565
Block CUDA graphs: decide before recording, release the pool, cap FSDP2 kept weights, drop unneeded copies by @danielhanchen in #15352
Studio: keep ticked check boxes when reading HTML files by @NilayYadav in #15601
Studio: prompt-length dims stay dynamic for positional block inputs under group offload hooks by @danielhanchen in #15653
Studio: vendor shadcn's tailwind.css instead of installing the shadcn CLI by @danielhanchen in #15635
Studio: honour context_overflow on the /v1/messages passthrough by @Shreesh-Sree in #15418
Studio: bound expires_in_days when creating an API key by @vineethsaivs in #14725
Studio Desktop: remove the duplicate KDE Wayland title bar on Linux by @Etherll in #15568
Studio diffusion: the pin budget credits clean mapped page cache inside a capped cgroup by @danielhanchen in #15642
Clarify training setup feedback and the music entry flow by @Imagineer99 in #15535
Studio: memoize per-repo lookups inside one model inventory scan by @danielhanchen in #15578
Studio: encode transient PNGs at zlib level 1 by @danielhanchen in #15579
Studio: tidy the System page Live monitor and GPU devices by @shimmyshimmer in #15705
Studio: find the backend's child processes from /proc on Linux for /api/system by @danielhanchen in #15592
MLX: build trl.all on first read instead of importing every TRL export at import by @danielhanchen in #15561
Studio: walk GGUF tokenizer string arrays through one buffer by @danielhanchen in #15593
Studio: resolve the gallery folder once per image and video listing by @danielhanchen in #15580
Studio: hash tokenizers for dataset cache names without their init-only vocab copies by @danielhanchen in #15577
Studio: answer slow cache sizes at once and walk cache trees faster by @danielhanchen in #15627
Studio: clear only the compiled cache unsloth_zoo writes to by @danielhanchen in #15628
Studio: launch full-access tool processes with start_new_session instead of a setsid pre-exec by @danielhanchen in #15595
CLI: defer YAML, asyncio, rich and the config schema build off import by @danielhanchen in #15585
Studio: bound tool-call stream scans and decode passthrough lines once by @danielhanchen in #15591
Studio: step streaming generators on a dedicated thread instead of the pool by @danielhanchen in #15590
Studio: decode and parse gpt-oss Harmony streams incrementally by @danielhanchen in #15589
Studio: check local single-file checkpoints without importing the diffusion stack by @danielhanchen in #15587
Studio: skip the passthrough healer reparse while no held call can close by @danielhanchen in #15618
Studio: put downloaded GGUFs under the right On Device filter by @NilayYadav in #15571
Studio: launch sandboxed tool processes without a Python pre-exec on Linux by @danielhanchen in #15625
Studio: load local JSON datasets whose later rows add keys by @danielhanchen in #15555
Studio: parse each seed glob and matched file once when applying loader skip rules by @danielhanchen in #15668
Studio: count ASCII in context estimates without a per-character loop by @danielhanchen in #15659
Studio: keep ticked PDF form check boxes in knowledge bases by @NilayYadav in #15572
Studio: faster MLX chat for models loaded in 4-bit by @NilayYadav in #15617
Studio: keep Qwen-Image-2.1 prompt-length dims dynamic under group offload hooks by @danielhanchen in #15652
Fix Online DPO crashing at start when use_vllm is on by @NilayYadav in #15598
Fix GRPO masking of short stop-string completions by @danielhanchen in #15704
Studio: only trust file search and dictation servers Studio started by @NilayYadav in #15600
Studio video: park MiniMax-H3's rotating VAEs by re-pointing at their pageable host copy when pinning is not allowed by @danielhanchen in #15715
[Bug] AMD: Unsloth Studio Images retries the ROCm sd.cpp build on hosts whose ROCm comes only from the torch wheel by @LeoBorcherding in #15430
Studio: reject tool calls from truncated local generations by @oobabooga in #15428
Studio: answer 422 to a timestamp or page cursor sqlite cannot store by @vineethsaivs in #15389
Native FSDP2: clip gradients through Unsloth's clip, not the mixed mesh path by @danielhanchen in #15708
fast_inference under UNSLOTH_FSDP2=1 torchrun: keep vLLM instead of refusing it by @danielhanchen in #15709
UNSLOTH_PP_SIZE: install the Trainer wrapper at import, so encoders refuse instead of training as DDP by @danielhanchen in #15710
Unsloth Studio / Desktop: log out, show the owner's profile in Accounts, and let the owner sign in with a chosen name by @LeoBorcherding in #15619
[bugfix] Snapshot in-memory DPO policies for the default reference by @taking-lying-flat in #15537
Studio: calmer login and setup pages with an accent sign-in button by @shimmyshimmer in #15731
Studio: clarify LoRA and training hyperparameter labels by @danielhanchen in #15479
Model loading: skip the root index probe when the repo listing shows no index by @danielhanchen in #15662
Studio: overlap model checkpoint downloads with dataset preparation by @danielhanchen in #15483
Studio desktop: give port validation probes the health watchdog's 10 s budget by @danielhanchen in #15670
Studio: copy generated SNAC codes to the host once per decode by @danielhanchen in #15666
Studio: reuse the size probe's repo listing for scoped download blob hashes by @danielhanchen in #15663
Studio: copy SNAC codes to the host once per example in TTS preprocessing by @danielhanchen in #15660
Studio: tokenize a repeated chunk text once in the sentence-transformers counter by @danielhanchen in #15664
Studio: match audio.cpp CUDA runtime to bundle architecture by @danielhanchen in #15491
Studio: read each PDF page's words once when locating chunk regions by @danielhanchen in #15665
Studio: convert only the cached unstructured seed rows a caller reads by @danielhanchen in #15667
Studio: answer ASCII tool result budgets without a per-character scan by @danielhanchen in #15676
Studio: render /openapi.json once per schema, off the event loop by @danielhanchen in #15684
Studio: match /assets before the API routes by @danielhanchen in #15669
Studio: read each dist-info METADATA once per install check by @danielhanchen in #15690
Studio: delete a RAG document's index rows in one scan per table by @danielhanchen in #15661
Studio: look up model picker rows by index map instead of a scan per row by @danielhanchen in #15674
Studio: mention ARM64 in Windows sandbox architecture guidance by @danielhanchen in #15494
Studio: stop re-stripping an open tool call body on every streamed token by @danielhanchen in #15707
Studio desktop: start the launch preflight's CLI -h probe during setup by @danielhanchen in #15671
Studio desktop: run -h and desktop-capabilities at once on a capability cache miss by @danielhanchen in #15672
Studio: visible light-mode borders, aligned media pickers, new image icons, modern login by @shimmyshimmer in #15726
Studio: count sd-cli WebM frames from VP8 packet headers in inspect_video by @danielhanchen in #15699
Studio: make Stop work while an MLX model is still reading the prompt by @NilayYadav in #15615
Studio desktop: share one loopback HTTP client per timeout by @danielhanchen in #15675
Studio: walk each GGUF tokenizer array once per process and seek past it on later reads by @danielhanchen in #15687
Faster import unsloth and from_pretrained: fewer GC passes, skip regexes that cannot match by @danielhanchen in #15560
Studio: log a live generation rate while a reply is still streaming by @NilayYadav in #15390
Studio desktop: leave the startup screen on server-port instead of at the next 500 ms poll by @danielhanchen in #15673
Studio: answer keyless admission on the event loop when nothing needs reading by @danielhanchen in #15683
Studio: report first-step training and gradient accumulation by @danielhanchen in #15480
Studio: acquire decision datasets alongside model loading by @danielhanchen in #15481
Studio: compose the FP8 and NVFP4 compile cache tags in a fixed order by @danielhanchen in #15692
Studio: precompile bytecode after setup so the first launch does not by @danielhanchen in #15646
Studio: advertise the chat-title repair guards on the thread lists instead of fetching /openapi.json by @danielhanchen in #15685
Keep an earlier model's smart offloading when a short sequence model loads by @danielhanchen in #15711
Run xformers attention on the tensors' device (device_map across GPUs gave NaN loss) by @danielhanchen in #15712
Studio diffusion: run the fused-dequant int8 GEMM on untuned sm80+ archs behind a probe and latency check by @danielhanchen in #15654
Studio: show cache sizes and clear individual Hugging Face model copies by @danielhanchen in #15488
Tests: bring three new tests in line with the port and encoding guards by @danielhanchen in #15745
Studio: queue /v1/videos while a video renders and batch waiting /v1/images/generations requests by @oobabooga in #15613
Studio training: log the traceback when a model fails to load by @debojyoti-dev in #15620
Studio: bound the passthrough healer gate to openers the parser would scan by @danielhanchen in #15717
setup.sh: keep the replaced-script handoff directly behind the dependency pass by @danielhanchen in #15747
Dependencies: drop four Python packages nothing imports (tyro, flatbuffers, pytorch_tokenizers, mammoth) by @danielhanchen in #15697
Studio: let the router skip routes whose path prefix cannot match by @danielhanchen in #15682
Studio: reuse VAE tile blend weights in tiled encode and LTX-2 tiled decode by @danielhanchen in #15700
Studio: refresh cache sizes that are being re-measured instead of showing the old number by @danielhanchen in #15718
Studio: web fetch tokenizes a page once and replays it to the later main-content passes by @danielhanchen in #15686
Studio: keep a Windows console stderr writable while the worker mirror owns fd 2 by @danielhanchen in #15734
Studio: Trick or treat sidebar row with Halloween themes and spooky greetings by @shimmyshimmer in #15678
Studio: read the VLM and audio detection sets from the transformers registry source instead of importing torch by @danielhanchen in #15688
Fix Gallery save toast to use the chosen filename by @Imagineer99 in #15725
Packed attention: full finetuning rows reach varlen (fp32 Q / K under autocast) by @danielhanchen in #15714
Studio: fewer GC passes while the backend imports, RSA key generated alongside startup by @danielhanchen in #15586
CLI: load command modules other than studio only when a command is dispatched by @danielhanchen in #15689
Add AGPLv3 header to desktop release workflow by @wasimysaid in #15755
Update OpenEnv links to openenv/OpenEnv by @sergiopaniego in #15743
Saving: stop repeated gc.collect passes once a pass finds nothing by @danielhanchen in #15694
Studio desktop: hide the window as soon as a quit is committed, and exit the backend without finalization by @Lyxot in #15723
Tests: follow the media picker's Chat-aligned height, start pwsh through run_pwsh by @danielhanchen in #15758
Studio: start the torch warm-up stages after the first response is sent by @danielhanchen in #15732
feat(studio): idle auto-unload an MLX chat model by @Lyxot in #15680
Studio: remove private helpers and constants that nothing uses by @danielhanchen in #15696
Studio: Search and Tools on by default for local models by @shimmyshimmer in #15644
Studio desktop: build release with fat LTO and one codegen unit by @danielhanchen in #15566
Studio: drop three frontend dependencies nothing imports by @danielhanchen in #15634
Studio desktop: trim the quit hide comments by @danielhanchen in #15760
Keep RAG WAL sidecars between database operations by @Imagineer99 in #15719
Studio diffusion: CUDA graph re-check fixes (eager OOM retry on originals, planned negative length records on the first render) by @danielhanchen in #14682
Studio diffusion: off-GIL CUDA prime does not latch before torch is imported by @danielhanchen in #14683
Studio FLUX.3 upsampler: pin the sidecar to the native sd.cpp image card by @danielhanchen in #14696
Studio video: LTX-2.5 dense text-encoder fallback room check prices the load dtype by @danielhanchen in #14664
Studio fine-tuning queue: run the batch model probe behind the account's path guard by @danielhanchen in #15727
Studio llama-server args: a value-less ngram-mod flag keeps the next flag by @danielhanchen in #15728
Tests: the two gc-skip auto dist tests take their port from the shared helper by @danielhanchen in #15729
Studio login: say Unsloth and size the title with the UI font scale by @danielhanchen in #15759
Studio: train ChatML datasets whose roles are capitalized by @NilayYadav in #15599
Studio frontend: bump zod to 4.6.7 (keep katex on 0.16) by @dependabot[bot] in #15746
Bump the pip group across 1 directory with 2 updates by @dependabot[bot] in #15518
Bump fast-uri from 3.1.8 to 3.1.9 in /docker/jupyter/unsloth_labext in the npm_and_yarn group across 1 directory by @dependabot[bot] in #15637
Bump the cargo-tauri group in /studio/src-tauri with 4 updates by @dependabot[bot] in #15360
Studio: confirm Full Access directly from a chat tool approval by @danielhanchen in #15485
CI: record the lazy CLI command import in the risky-loader baseline by @danielhanchen in #15762
AppImage E2E: count the app shell as rendered when the splash is already gone by @danielhanchen in #15764
FSDP2 block CUDA graphs: free the gathered weights of MoE blocks by @danielhanchen in #15720
cuDNN SDPA grad layout fix: let compiled code modify the SDPA output in place by @danielhanchen in #15721
Studio: speedometer icon for generation speed on the API page by @shimmyshimmer in #15769
Studio: spider or bats on a third of Halloween pages each by @shimmyshimmer in #15770
Studio: cast-cache writer reads the host copy offload already keeps instead of waiting out an onload by @danielhanchen in #15698
Studio: preview JSON Lines datasets with the stdlib, not pyarrow's JSON reader by @lonexreb in #15713
Clarify Deep Research handoff for research briefs by @Imagineer99 in #15724
Studio: batch packed fine-tuning jobs, each job packing only its own documents by @danielhanchen in #15733
Keep packed documents apart under flash attention when layers get mRoPE position ids by @danielhanchen in #15771
Installer downloads: retry a body cut short or trickling instead of hashing it by @danielhanchen in #15763
Studio: read the Int8Tensor constructor signature once per class in the native INT8 rebuild by @danielhanchen in #15701
Studio: load the attachment hasher before deleting a chat attachment by @danielhanchen in #15751
Studio: decode each relayed SSE line once in the tool call stripper by @danielhanchen in #15753
Studio: hash the tokenizer fingerprint without touching the live tokenizer by @danielhanchen in #15749
Studio: skip the DFlash size request when no sidecar is eligible by @danielhanchen in #15754
Studio video gallery: clear() reads and unlinks sidecars in the folder it locked by @danielhanchen in #15766
Studio: install the released Diffusers 0.42.1 instead of the pinned main commit by @danielhanchen in #15761
Tests: follow the thread-list guard flag and the RAG WAL keeper by @danielhanchen in #15778
Import: check for a broken openai install on first use instead of at import unsloth by @danielhanchen in #15702

New Contributors

@Aafimalek made their first contribution in #15421
@Shreesh-Sree made their first contribution in #15418
@sergiopaniego made their first contribution in #15743

Full Changelog: v0.1.911-beta...v0.1.912-beta

unslothai/unsloth v0.1.912-beta Trick or Treat + Optimizations on GitHub