v0.1.907-beta

Unsloth has gotten major performance upgrades!

## Highlights

- **Large perf gains**: all Unsloth UI/UX faster - up to 10x for multiple chats
- **Up to 5x faster inference**: 95% cache hit rate, MoE Strata style fixes, slowdown fixes
- **Better Compaction** - works for all API + better UI handling
- **Drag and drop files**: Desktop works again with more than 3× more supported files
- **Browser use with scripts**: click, move the mouse, drag, run approved scripts and set per-site browser rules.
- **LTX-2.5 + Wan2.2 + Faster image LoRA training**: new video generation, image-to-video and faster image training.
- **NVIDIA RTX Spark**: now supported with our Windows ARM64 downloadables
- **Major Unsloth workflow upgrades**: improved files, chat history, browser agent and document handling.
- **Improved hardware compatibility**: Windows AMD, Intel XPU and CUDA/PyTorch fixes.
- **Faster image LoRA training** with aspect buckets, fused AdamW and improved FP8/MXFP8 support.

## Performance

- Large chat and training histories load in smaller batches, using less startup data and memory.
- Follow-up replies start sooner in long GGUF chats, including conversations with many tool calls.
- LTX-2.5 generation runs faster with the default two-stage settings when memory permits.
- Qwen3.5, Qwen3.6 and Qwen3.8 get faster prefill and speculative decoding with MLX.

## Browser use

- Models can click at a point, move the mouse and drag page elements, including unlabelled targets such as canvases.
- Run scripts on a page with your approval each time, or set Run scripts to Never in Settings > Browser.
- Set per-site rules for unlabelled clicks, including Always allow for a site you trust.
- Better handling of compressed pages, client-side navigation and delayed content in the web UI. Models can also fill date and time fields.
- Pressing Enter in a form gets the same approval check as clicking its submit button.

## Images + video

- Generate videos with LTX-2.5 from text or a start image, with synchronized audio. [GGUF](https://huggingface.co/unsloth/LTX-2.5-GGUF)
- Add a Start frame with Wan2.2-TI2V-5B, LTX-2 or LTX-2.3 to animate an image on the same loaded model.
- Choose SDXL samplers and schedules on the Images page, and load v-prediction checkpoints correctly.
- Crop image inputs before generation. Large photo uploads and ordinary video frame uploads resize automatically.

## Training

- Image LoRA training uses fused AdamW by default on NVIDIA GPUs, with aspect-ratio buckets for new supported runs.
- FP8 LoRA training for FLUX and Z-Image uses less memory on NVIDIA Ada, Hopper and Blackwell GPUs. MXFP8 handles aspect buckets on Blackwell with a compatible torchao stack.
- Set Sample every on the Train tab to see preview images during supported image training and in run history.
- Set caption dropout, LoRA dropout and checkpoint retention on the image Train tab. Random flipping is now opt-in.
- Fine-tune AWQ models with LoRA using Transformers 5 and GPTQModel, and export INT4 QAT models as quantized weights.

## Chat + Desktop

- Drag files into Desktop chat again, including more Office documents, templates, EPUBs, README files, dotfiles and media.
- Upload supported PDF and Office documents up to 200 MB in the web UI.
- Linked folders honor common `.gitignore` rules. Upload folders to project Sources from your browser, retry failed uploads and open completed sources.
- Type `/rewind`, or choose Rewind from a message's More menu, to restore the conversation, tracked files in the chat folder, or both. Preview changes and undo a restore.
- Use `/compact` in saved GGUF or MLX chats that support tools to compact context on the next message. Inline dividers mark compaction.
- Download model-created files from tool results, and enable notifications when a background reply finishes.

## Models + settings

- Use local models in VS Code Copilot Chat with `unsloth start vscode`.
- Coding agents can use Unsloth over MCP with an account API key to load models, generate, train and export.
- Choose a model download folder, and reveal or delete listed models from LM Studio, Ollama and your own folders.

## Hardware compatibility

- Fix the PyTorch/Triton startup error on Windows AMD systems.
- Intel XPU recognizes native BF16 on supported Arc and Data Center GPUs without intel-ocloc, with improved GPU selection on single-tile Linux hosts.
- Avoid stale Triton launcher errors after switching Python versions on CUDA systems.
- Improved low-VRAM Gemma 4 E2B loading, including host-memory offload on Windows and WSL.

## What's Changed
* Bump install.sh / install.ps1 pins to unsloth>=2026.10.4 by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13433
* Key triton < 3.4 driver module cache by Python version by @danielhanchen in https://github.com/unslothai/unsloth/pull/13409
* Studio: linked folders honour .gitignore, folder uploads in the browser, queued drops and retry for failed sources by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13224
* Studio: image-to-video for Wan2.2-TI2V-5B and LTX-2 / LTX-2.3 by @danielhanchen in https://github.com/unslothai/unsloth/pull/13378
* fix(studio): preserve chat search position on close by @Sneakr in https://github.com/unslothai/unsloth/pull/13437
* Studio: refuse a text-encoder GGUF picked as the image or video model by @danielhanchen in https://github.com/unslothai/unsloth/pull/13423
* Studio: keep emoji split across tokens in MLX replies by @danielhanchen in https://github.com/unslothai/unsloth/pull/13416
* Studio: retry from the Hub when a cached snapshot has no tokenizer files by @danielhanchen in https://github.com/unslothai/unsloth/pull/13412
* Unsloth Studio (AMD/Windows): import PyTorch before Triton to avoid WinError 126 by @boobcactus in https://github.com/unslothai/unsloth/pull/10932
* Studio: list each on-disk quant of a multi-quant GGUF in GET /v1/models by @danielhanchen in https://github.com/unslothai/unsloth/pull/13422
* Studio: split DiffusionGemma across GPUs when one GPU cannot hold it by @danielhanchen in https://github.com/unslothai/unsloth/pull/13291
* Tests: catch up with four product changes that landed red on main by @danielhanchen in https://github.com/unslothai/unsloth/pull/13445
* fix(studio): show queued downloads consistently across media routes by @Imagineer99 in https://github.com/unslothai/unsloth/pull/13219
* Accept bf16 tie-flips in the batched left-padded generation guard by @danielhanchen in https://github.com/unslothai/unsloth/pull/13408
* fix: reshape noncontiguous gradients in fused LoRA backward by @MrCapricornLiu in https://github.com/unslothai/unsloth/pull/10712
* test_trl_cap_sites: unsloth_zoo 2026.10.4 admits trl 1.15.0, end the deferral by @danielhanchen in https://github.com/unslothai/unsloth/pull/13449
* Studio: keep Spark-TTS LoRA checkpoints exportable by @danielhanchen in https://github.com/unslothai/unsloth/pull/13424
* Keep uv, its cache and the uv-managed Python inside a custom UNSLOTH_STUDIO_HOME by @danielhanchen in https://github.com/unslothai/unsloth/pull/13373
* Studio: browser use follow-ups (safetensors nudge, date fields, disabled dropdowns, moved pages) by @danielhanchen in https://github.com/unslothai/unsloth/pull/13429
* Keep a user's UNSLOTH_RETURN_LOGITS=1 after for_inference() on FastModel by @danielhanchen in https://github.com/unslothai/unsloth/pull/13406
* Codex device-poll tests: fake asyncio.sleep only on the test's own thread by @danielhanchen in https://github.com/unslothai/unsloth/pull/13457
* Format two test files that #13378 and #13437 left off the formatter's fixed point by @danielhanchen in https://github.com/unslothai/unsloth/pull/13462
* Studio: run LTX-2.5 on the Video page by @danielhanchen in https://github.com/unslothai/unsloth/pull/13463
* Baseline the unsloth-zoo 2026.10.4 and trl 1.15.0 findings after review by @danielhanchen in https://github.com/unslothai/unsloth/pull/13472
* studio: restore file drops and composer drag feedback by @mahiatlinux in https://github.com/unslothai/unsloth/pull/13387
* Studio: optically centre the annotate button's icon by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13495
* Fix the three Backend CI reds LTX-2.5 left on main by @danielhanchen in https://github.com/unslothai/unsloth/pull/13496
* Studio: Open WebUI llama.cpp residency and explicit load/unload (#10610) by @debojyoti-dev in https://github.com/unslothai/unsloth/pull/10693
* studio: keep open thinking sized when the chat pane narrows by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13503
* Studio: nudge the dictate mic down 1px instead of 2px by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13504
* Studio: larger Browser use icon, more room and a shorter description in Settings by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13506
* tests: give each xdist worker one CPU share for torch by @danielhanchen in https://github.com/unslothai/unsloth/pull/13494
* Studio: close the LTX-2.5 two-stage speed gaps in a shared multi-stage layer by @danielhanchen in https://github.com/unslothai/unsloth/pull/13508
* Update README.md by @danielhanchen in https://github.com/unslothai/unsloth/pull/13512
* Studio: clip the browser page at the split handle while dragging it smaller by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13492
* Studio: composer accepts every document type chat with files parses by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13426
* studio: stronger light-mode Library card hover, centered time pill by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13509
* Studio: composer takes README-style files, dotfiles, Office templates and more media by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13493
* Studio: LTX-2.5 VAE fixture, catalog row and Video page defaults by @danielhanchen in https://github.com/unslothai/unsloth/pull/13500
* Make UNSLOTH_STUDIO_KEEP_LEGACY=0 work through docker/run.sh by @NilayYadav in https://github.com/unslothai/unsloth/pull/13488
* Studio: add --models-dir to unsloth studio for local model folders by @danielhanchen in https://github.com/unslothai/unsloth/pull/13435
* Studio: seed LTX-2.5 fp8 / int8 loads from hosted pre-quantized DiTs by @danielhanchen in https://github.com/unslothai/unsloth/pull/13516
* Studio: never follow a symlink when linking a file into a tool workspace by @danielhanchen in https://github.com/unslothai/unsloth/pull/13513
* studio: dim selected Library cards, slightly tighter card corners by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13518
* Studio: offer a download for files edit_file creates by @danielhanchen in https://github.com/unslothai/unsloth/pull/13448
* Studio: diffusion auto-optimizer core shared by the image and video load paths by @danielhanchen in https://github.com/unslothai/unsloth/pull/13520
* Studio: Regenerate title only in the Library's chat menu by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13491
* Add unsloth start vscode: put the served model in VS Code's Copilot Chat model picker by @danielhanchen in https://github.com/unslothai/unsloth/pull/13439
* Studio: mark where a chat was compacted with an inline divider by @oobabooga in https://github.com/unslothai/unsloth/pull/13502
* Studio: serve HTTPS from your own certificate (--ssl-certfile / --ssl-keyfile) by @danielhanchen in https://github.com/unslothai/unsloth/pull/13447
* Studio: show pending llama.cpp, whisper.cpp and audio.cpp updates on one card by @oobabooga in https://github.com/unslothai/unsloth/pull/13497
* Studio: open a project source from the Sources tab by @danielhanchen in https://github.com/unslothai/unsloth/pull/13441
* Studio: diffusion flow_shift is the effective shift, and min-SNR is recorded as unused for flow-matching runs by @danielhanchen in https://github.com/unslothai/unsloth/pull/13215
* Studio: random flip opt-in, plus checkpoint retention, caption dropout, LoRA dropout and timestep shift on the image Train tab by @danielhanchen in https://github.com/unslothai/unsloth/pull/13210
* Studio: scale the SDXL loss in fp16 and keep image LoRA training off devices it cannot use by @danielhanchen in https://github.com/unslothai/unsloth/pull/13278
* Studio: train image LoRAs with aspect-ratio buckets by @danielhanchen in https://github.com/unslothai/unsloth/pull/13218
* Format the diffusion auto-optimizer and LTX-2.5 prequant files by @danielhanchen in https://github.com/unslothai/unsloth/pull/13526
* Keep the saved Hugging Face login when from_pretrained gets a token by @NilayYadav in https://github.com/unslothai/unsloth/pull/13486
* test_packing_sdpa: take the loss path it reads labels from by @danielhanchen in https://github.com/unslothai/unsloth/pull/13434
* Studio: sample images during image LoRA training by @danielhanchen in https://github.com/unslothai/unsloth/pull/13220
* Turn off Granite 3.0-3.2 attention dropout on float16 loads by @danielhanchen in https://github.com/unslothai/unsloth/pull/13417
* Studio: train Z-Image LoRAs on attention and feed-forward by @danielhanchen in https://github.com/unslothai/unsloth/pull/13233
* Studio installer (AMD, Linux): no generic ROCm torch for RDNA 1 / Polaris when the arch is readable by @danielhanchen in https://github.com/unslothai/unsloth/pull/13419
* Studio: reveal and delete for LM Studio, Ollama, custom-folder and other local models by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13507
* Re-measure the Studio startup budget after the Library and composer merges by @danielhanchen in https://github.com/unslothai/unsloth/pull/13533
* Studio: per-family diffusion tables move into the family registries by @danielhanchen in https://github.com/unslothai/unsloth/pull/13536
* Studio: run the Shutdown cleanup off the event loop by @danielhanchen in https://github.com/unslothai/unsloth/pull/13534
* Save int4 QAT LoRA models as real int4 with save_pretrained_torchao by @danielhanchen in https://github.com/unslothai/unsloth/pull/13413
* Studio (AMD iGPU): split an iGPU row's total into VRAM and shared memory by @danielhanchen in https://github.com/unslothai/unsloth/pull/13421
* Fix the exec-literal, encoding and formatter gates left red on main by @danielhanchen in https://github.com/unslothai/unsloth/pull/13543
* Studio: transcribe with downloaded MLX speech-to-text models on Apple Silicon by @danielhanchen in https://github.com/unslothai/unsloth/pull/13443
* Studio: cap Qwen3.8 NPU context at 8192 by @oobabooga in https://github.com/unslothai/unsloth/pull/13525
* Studio: XPU native bf16 no longer depends on intel-ocloc by @danielhanchen in https://github.com/unslothai/unsloth/pull/13527
* Studio: rename the composer's Code toggle to Tools by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13402
* Studio: rounded-rectangle hover on the header model selector by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13546
* Studio: let a model loaded at a set context reload on Auto by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13555
* Studio: read ZE_AFFINITY_MASK the way Level Zero does by @danielhanchen in https://github.com/unslothai/unsloth/pull/13528
* unsloth start opencode: detect OpenCode V2 installed as plain opencode by @danielhanchen in https://github.com/unslothai/unsloth/pull/13550
* Studio: size music waits for Metal and Vulkan, not only CPU by @danielhanchen in https://github.com/unslothai/unsloth/pull/13551
* Skip bitsandbytes' per-parameter device sync in non-paged optimizer steps by @danielhanchen in https://github.com/unslothai/unsloth/pull/13389
* Studio: explicit XPU GPU ids on single-tile Intel hosts by @danielhanchen in https://github.com/unslothai/unsloth/pull/13540
* Studio: keep Hub Run settings open on a first Chat visit by @oobabooga in https://github.com/unslothai/unsloth/pull/13524
* Studio: run AutoRound checkpoints on vLLM, and tidy the engine dialog by @danielhanchen in https://github.com/unslothai/unsloth/pull/13405
* fix(studio): handle harmless Node helper stderr on Windows by @Imagineer99 in https://github.com/unslothai/unsloth/pull/13530
* Studio: faster image LoRA training (fused AdamW, compile nf4 bases) by @danielhanchen in https://github.com/unslothai/unsloth/pull/13236
* Studio: run LoftQ when the LoftQ variant is selected by @danielhanchen in https://github.com/unslothai/unsloth/pull/13535
* Fix int8-int4 QAT LoRA at low ranks and on tied-embedding models by @danielhanchen in https://github.com/unslothai/unsloth/pull/13552
* Studio: skip a DFlash drafter that an image would break under Auto by @danielhanchen in https://github.com/unslothai/unsloth/pull/13553
* Studio: stop LTX-2 loads crashing Windows with an access violation on 32 GB machines by @danielhanchen in https://github.com/unslothai/unsloth/pull/13432
* Studio: per-chat file history and /rewind by @danielhanchen in https://github.com/unslothai/unsloth/pull/13522
* Studio: browser agent can click at a point, move the mouse, drag, and run approved scripts by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13505
* Studio: list VS Code and Pi as coding agents and reorder the agent picker by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13547
* Studio: Regenerate title in the chat header menu by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13574
* Studio: rename sidebar chats in the Rename chat dialog by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13575
* Studio: translate the chat welcome greeting in every locale by @danielhanchen in https://github.com/unslothai/unsloth/pull/13549
* Studio: Run scripts setting and per-site Unlabelled clicks rule for the browser agent by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13529
* Studio: fp8 / mxfp8 image LoRA training on every aspect bucket, mxfp8 for biased linears, fp8 sm89 floor by @danielhanchen in https://github.com/unslothai/unsloth/pull/13317
* Re-measure the Studio startup budget after the file history and agent merges by @danielhanchen in https://github.com/unslothai/unsloth/pull/13586
* Studio: optional notification when a chat reply finishes in the background by @danielhanchen in https://github.com/unslothai/unsloth/pull/13554
* Prebuilt CUDA wheels: torch 2.11 for Windows x64 by @danielhanchen in https://github.com/unslothai/unsloth/pull/13563
* Studio: fix compressed pages, router paths and early empty reads in the browser panel by @oobabooga in https://github.com/unslothai/unsloth/pull/13532
* Studio: popups dim without blur, with a faint shadow; Settings a step darker by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13577
* Keep torch's bundled flash on Hopper for narrow models under UNSLOTH_FLASH_ATTENTION_IMPL=auto by @danielhanchen in https://github.com/unslothai/unsloth/pull/13544
* Parse each forward once in the per-step loss-kwargs walk by @danielhanchen in https://github.com/unslothai/unsloth/pull/13585
* perf(studio): enter Zoo's exact verify, NAX and packed gated-delta scopes on Unsloth Studio's own MLX paths by @Lyxot in https://github.com/unslothai/unsloth/pull/13404
* Fit Gemma 4 E2B on 8 GB cards on Windows and in Studio chat by @danielhanchen in https://github.com/unslothai/unsloth/pull/13425
* Let AWQ checkpoints train with LoRA by @danielhanchen in https://github.com/unslothai/unsloth/pull/13584
* Pass str target_modules (regex or "all-linear") to PEFT on the Llama path by @danielhanchen in https://github.com/unslothai/unsloth/pull/13567
* Patch TRL's served-teacher, GOLD and SDFT distillation trainers by @danielhanchen in https://github.com/unslothai/unsloth/pull/13580
* Studio: downscale oversized image uploads and add crop to image inputs by @danielhanchen in https://github.com/unslothai/unsloth/pull/13561
* Run non-LoRA PEFT adapters through their own forward in fast inference by @danielhanchen in https://github.com/unslothai/unsloth/pull/13578
* Make the weights_at_root call-site and diffusion autoopt parity tests host independent by @danielhanchen in https://github.com/unslothai/unsloth/pull/13582
* studiobench UI parity: report a UI change the PR declares instead of failing by @danielhanchen in https://github.com/unslothai/unsloth/pull/13583
* Studio setup: build the frontend in the background on source installs (setup.sh and setup.ps1) by @danielhanchen in https://github.com/unslothai/unsloth/pull/13442
* Studio: load sidebar chats, projects and training runs 20 at a time from the server by @danielhanchen in https://github.com/unslothai/unsloth/pull/13420
* Fix image inputs in Clef training and prediction by @Imagineer99 in https://github.com/unslothai/unsloth/pull/13471
* Studio: apply the tool output length setting to MCP tool results by @danielhanchen in https://github.com/unslothai/unsloth/pull/13588
* Tests: follow the sidebar rename dialog and run the H3 compile-order test without diffusers by @danielhanchen in https://github.com/unslothai/unsloth/pull/13590
* Studio: let coding agents use Unsloth Studio over MCP with API keys by @Etherll in https://github.com/unslothai/unsloth/pull/13118
* Studio: pick the sampler on the Images page and load v-prediction SDXL checkpoints correctly by @danielhanchen in https://github.com/unslothai/unsloth/pull/13569
* Studio: add a Download to folder for model downloads by @danielhanchen in https://github.com/unslothai/unsloth/pull/13557
* Studio: keep the Transcribe loaded state and Eject in step with the sidecar by @Etherll in https://github.com/unslothai/unsloth/pull/13514
* Studio: fix Spark-TTS training failing on None/config.yaml (BiCodec repo root unset) by @carlitogwapito in https://github.com/unslothai/unsloth/pull/13465
* Studio: add /compact and draw compactions where they happened by @danielhanchen in https://github.com/unslothai/unsloth/pull/13562
* Studio: keep the prompt cache across long tool chats and follow-up turns by @danielhanchen in https://github.com/unslothai/unsloth/pull/13456
* Keep Mistral Small 3.2 tool definitions in training rows by @danielhanchen in https://github.com/unslothai/unsloth/pull/13597
* Studio: video component sizes from safetensors headers, measured peaks for every video family by @danielhanchen in https://github.com/unslothai/unsloth/pull/13537
* Studio: fix audio.cpp partial downloads, aligner progress, CPU fallback by @Etherll in https://github.com/unslothai/unsloth/pull/13515
* Studio: faster browser agent page reads after a load, and leaner drags by @danielhanchen in https://github.com/unslothai/unsloth/pull/13602
* Studio: trim comments in the video header sizes code by @danielhanchen in https://github.com/unslothai/unsloth/pull/13610
* Studio: optional Chart view with labelled axes for weekly and cumulative token activity by @danielhanchen in https://github.com/unslothai/unsloth/pull/13558
* Studio: fit MiniMax-H3 GGUF to a Vulkan host's GTT pool before rendering by @danielhanchen in https://github.com/unslothai/unsloth/pull/13531
* Studio: LTX-2.5 GGUF quants (unsloth/LTX-2.5-GGUF) by @danielhanchen in https://github.com/unslothai/unsloth/pull/13573
* FastSentenceTransformer: reuse the loaded model for Transformer subclass modules by @danielhanchen in https://github.com/unslothai/unsloth/pull/13591
* Studio: list, reveal and delete audio.cpp umbrella models by folder by @Etherll in https://github.com/unslothai/unsloth/pull/13519
* Fix the diffusion scheduler, parity, formatter and bundle reds on main by @danielhanchen in https://github.com/unslothai/unsloth/pull/13612
* Studio: Auto drops serving slots before the MTP drafter by @danielhanchen in https://github.com/unslothai/unsloth/pull/13455
* tests: resolve every closure name in the restoreSettings sandbox by @danielhanchen in https://github.com/unslothai/unsloth/pull/13617
* Studio: inline message actions, matching response editor by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13615
* Studio: set thinking traces apart from the answer and give tool rows kind icons by @shimmyshimmer in https://github.com/unslothai/unsloth/pull/13545
* Studio: fix the audio model selector's list, labels and search by @Etherll in https://github.com/unslothai/unsloth/pull/13517
* Studio tests: restore stub, titlebar backdrop, and run the autoopt parity fixture with diffusers by @danielhanchen in https://github.com/unslothai/unsloth/pull/13608
* Studio: emit --moe-cache-mib auto only where the cache measured faster, price the spill micro-batch by @danielhanchen in https://github.com/unslothai/unsloth/pull/13600

## New Contributors
* @boobcactus made their first contribution in https://github.com/unslothai/unsloth/pull/10932
* @carlitogwapito made their first contribution in https://github.com/unslothai/unsloth/pull/13465

**Full Changelog**: https://github.com/unslothai/unsloth/compare/v0.1.906-beta...v0.1.907-beta