{"id":11446,"date":"2026-11-17T11:00:00","date_gmt":"2026-11-17T02:00:00","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=11446"},"modified":"2026-11-17T11:09:19","modified_gmt":"2026-11-17T02:09:19","slug":"unsloth-v0-1-907-beta-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-907-beta-released\/","title":{"rendered":"Unsloth v0.1.907-beta Released with Major Speed and UI Upgrades"},"content":{"rendered":"<p><!-- lmw:facts --><\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Repository<\/td>\n<td><a href=\"https:\/\/github.com\/unslothai\/unsloth\">unslothai\/unsloth<\/a><\/td>\n<\/tr>\n<tr>\n<td>Version<\/td>\n<td><a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/tag\/v0.1.907-beta\">v0.1.907-beta<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-14<\/td>\n<\/tr>\n<tr>\n<td>License<\/td>\n<td>Apache-2.0<\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>Unsloth, a UI and toolkit for running and training large language models (LLMs) and diffusion models locally, has released its latest version, &#8220;v0.1.907-beta&#8221;. Unsloth is developed in Python and provided under the Apache-2.0 license.<\/p>\n<p>This version delivers major performance enhancements, including up to a 10x improvement in overall UI responsiveness, a 95% cache hit rate, and up to a 5x speedup in inference speeds thanks to fixes related to MoE Strata. Furthermore, it incorporates a wide range of new features such as expanded browser agent capabilities, support for new video generation models, and streamlined image LoRA training, making this an update well worth installing for users seeking faster performance in both inference and training.<\/p>\n<h2>Breaking Changes and Deprecations<\/h2>\n<p>There is one change in image training settings where the default behavior differs from previous versions.<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th style=\"text-align: left;\">Setting<\/th>\n<th style=\"text-align: left;\">Previous Behavior<\/th>\n<th style=\"text-align: left;\">Updated Behavior<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"text-align: left;\">Random flipping<\/td>\n<td style=\"text-align: left;\">Applied automatically<\/td>\n<td style=\"text-align: left;\">Opt-in (requires manual explicit activation)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>This change affects users running image LoRA training in Unsloth&#8217;s &#8220;Train&#8221; tab. If you want to use horizontal image flipping as data augmentation during training, you must manually enable the setting after updating.<\/p>\n<h2>Key Changes<\/h2>\n<h3>Significant Performance Improvements in Inference and UI Responsiveness<\/h3>\n<p>Overall UI operations have been accelerated, achieving up to a 10x speedup in scenarios handling multiple chats. In inference processing as well, performance gains of up to 5x have been confirmed, driven by achieving a 95% cache hit rate and MoE Strata-style fixes. Additionally, loading large chat and training histories has been changed to batch processing in smaller chunks, reducing data transfer volumes and memory consumption at startup. Response initiation times are now faster for long GGUF chats and conversations heavily utilizing tool calls. On Apple Silicon MLX environments, prefill processing and speculative decoding have been accelerated for Qwen3.5, Qwen3.6, and Qwen3.8.<\/p>\n<h3>Expansion of Browser Agent Features<\/h3>\n<p>Models can now click specified coordinates, move the mouse, and drag elements on a page\u2014including unlabeled elements such as canvases. Security has also been enhanced with settings to prompt the user for approval each time a script is executed on a page, or to set script execution to &#8220;Never&#8221; via the settings screen (Settings &gt; Browser). Domain-specific rules can also be configured, such as always allowing clicks on unlabeled elements for trusted sites. On the Web UI, tracking reliability has improved for lazy-loaded content, client-side transitions, and compressed pages, alongside support for auto-filling date and time fields.<\/p>\n<h3>Support for New Video Generation Models and Media Features<\/h3>\n<p>Support has been added for the video generation model &#8220;LTX-2.5&#8221;, enabling audio-inclusive video generation from text or start images. Additionally, Image-to-Video generation is now supported in &#8220;Wan2.2-TI2V-5B&#8221;, &#8220;LTX-2&#8221;, and &#8220;LTX-2.3&#8221; by adding a start frame on the same pre-loaded model to animate images. On the image generation side, additions include sampler and schedule options for SDXL, proper loading of v-prediction checkpoints, pre-generation image cropping, and automatic resizing for large photos and video frames.<\/p>\n<h3>Accelerated Image LoRA Training and New Features<\/h3>\n<p>Fused AdamW is now used by default for image LoRA training in NVIDIA environments, improving training efficiency alongside support for aspect ratio buckets. Memory consumption for FLUX and Z-Image FP8 LoRA training has been reduced on NVIDIA Ada, Hopper, and Blackwell environments, and Blackwell also supports aspect ratio bucket training for MXFP8 using a compatible torchao stack. In addition, new additions include a &#8220;Sample every&#8221; setting to generate preview images on the training tab, caption and LoRA dropout rate settings, and checkpoint retention settings. Furthermore, support has been added for LoRA fine-tuning of AWQ models using Transformers 5 and GPTQModel, as well as exporting INT4 QAT models as quantized weights.<\/p>\n<h3>Enhancements to Chat, Desktop, and Integration Features<\/h3>\n<p>File drag-and-drop in the desktop app has been restored, and supported formats\u2014including Office documents, EPUB, READMEs, and dotfiles\u2014have expanded more than threefold. In the Web UI, users can upload PDFs and Office documents up to 200MB. Chat functions now include a <code>\/rewind<\/code> command to restore conversations and working files, and a <code>\/compact<\/code> command to compress context. For external integrations, the <code>unsloth start vscode<\/code> command now registers local models in VS Code&#8217;s Copilot Chat model selection menu, and the model&#8217;s loading, generation, training, and export can now be controlled from coding agents via MCP (Model Context Protocol).<\/p>\n<h3>Hardware Compatibility Improvements and Bug Fixes<\/h3>\n<p>PyTorch and Triton startup errors that occurred in Windows AMD environments have been resolved. On Intel XPU environments, native BF16 is now recognized on supported Arc and datacenter GPUs without needing to install intel-ocloc, and GPU selection logic on single-tile Linux hosts has also been improved. Furthermore, errors in the Triton launcher occurring when switching Python versions in CUDA environments are now avoided, and stability when loading the Gemma 4 E2B model in low VRAM environments (including host memory offloading on Windows and WSL) has been improved.<\/p>\n<h2>Supported Models and Hardware<\/h2>\n<p>This update introduces numerous features, including support for the latest video generation models and optimizations across diverse hardware environments.<\/p>\n<h3>Video Generation and Image Model Support<\/h3>\n<ul>\n<li><strong>LTX-2.5<\/strong>: Added support for video generation from text or images. Inference in GGUF format is also possible, running fast with default two-stage settings as long as memory permits.<\/li>\n<li><strong>Wan2.2-TI2V-5B \/ LTX-2 \/ LTX-2.3<\/strong>: Enabled the Image-to-Video feature to animate images by adding a start frame on the same pre-loaded model.<\/li>\n<li><strong>Qwen3.5 \/ Qwen3.6 \/ Qwen3.8<\/strong>: Prefill processing and speculative decoding have been accelerated on Apple Silicon environments (MLX).<\/li>\n<li><strong>Gemma 4 E2B<\/strong>: Enabled operation even in low VRAM environments like 8GB by utilizing host memory offloading in Windows and WSL environments.<\/li>\n<li><strong>SDXL<\/strong>: Sampler and schedule selection are now available on the image generation page, and v-prediction checkpoints can now be loaded correctly.<\/li>\n<li><strong>AWQ \/ INT4 QAT<\/strong>: Added support for LoRA fine-tuning of AWQ models and exporting INT4 QAT models as quantized weights.<\/li>\n<li><strong>Speech and Music Models<\/strong>: Enabled speech-to-text using MLX on Apple Silicon, alongside support for audio and music generation via audio.cpp.<\/li>\n<\/ul>\n<h3>Expansion of Hardware and Accelerators<\/h3>\n<ul>\n<li><strong>NVIDIA RTX Spark<\/strong>: Newly supported in the download version for Windows ARM64.<\/li>\n<li><strong>NVIDIA Blackwell \/ Hopper \/ Ada<\/strong>: Memory consumption reduced for FLUX and Z-Image FP8 LoRA training. Blackwell generation specifically also allows aspect ratio bucket training in MXFP8 format.<\/li>\n<li><strong>Intel XPU<\/strong>: Native BF16 is now recognized on Arc and datacenter GPUs without requiring individual installation of intel-ocloc. GPU selection on single-tile Linux hosts has also been improved.<\/li>\n<li><strong>Windows AMD<\/strong>: Startup errors involving PyTorch and Triton have been fixed to improve stability. Detection of RDNA 1 and Polaris architectures has also been improved.<\/li>\n<\/ul>\n<h2>How to Get It<\/h2>\n<p>To update an existing environment to the latest v0.1.907-beta, run the following procedures.<\/p>\n<h3>For macOS, Linux, and WSL<\/h3>\n<p>Navigate to the repository directory in your terminal, fetch the latest sources, and re-run the installation script.<\/p>\n<pre><code class=\"language-bash\">cd unsloth &amp;&amp; git pull\n.\/install.sh --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<h3>For Windows (PowerShell)<\/h3>\n<p>Open PowerShell and similarly update the repository and run the installation script.<\/p>\n<pre><code class=\"language-powershell\">cd unsloth; git pull\n.\\\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<h3>If Uninstallation is Required<\/h3>\n<p>If you want to completely remove and rebuild your environment from scratch, you can use the official uninstallation scripts.<\/p>\n<p><strong>macOS, WSL, Linux:<\/strong><\/p>\n<pre><code class=\"language-bash\">curl -fsSL https:\/\/raw.githubusercontent.com\/unslothai\/unsloth\/main\/scripts\/uninstall.sh | sh\n<\/code><\/pre>\n<p><strong>Windows (PowerShell):<\/strong><\/p>\n<pre><code class=\"language-powershell\">irm https:\/\/raw.githubusercontent.com\/unslothai\/unsloth\/main\/scripts\/uninstall.ps1 | iex\n<\/code><\/pre>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-918-beta-release\/\">Unsloth v0.1.918-beta Adds Computer Use, MCP Game Engine Support<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-917-beta-released\/\">Unsloth v0.1.917-beta Released with Major Performance Boosts<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-916-beta-released\/\">Unsloth v0.1.916-beta Released: Performance &amp; Multi-GPU Updates<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-915-beta-released\/\">Unsloth v0.1.915-beta Released with Browser Use &amp; Qwen3.8<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>Follow this tool<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-unsloth-en\/\">Unsloth overview and release history (89 releases tracked)<\/a><\/li>\n<li><strong>Other quantization, model formats and fine-tuning<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-ggml-en\/\">ggml<\/a> \/ <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/engine-exllamav3-en\/\">ExLlamaV3<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/tag\/v0.1.907-beta\">https:\/\/github.com\/unslothai\/unsloth\/releases\/tag\/v0.1.907-beta<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13433\">https:\/\/github.com\/unslothai\/unsloth\/pull\/13433<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13505\">https:\/\/github.com\/unslothai\/unsloth\/pull\/13505<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13439\">https:\/\/github.com\/unslothai\/unsloth\/pull\/13439<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13527\">https:\/\/github.com\/unslothai\/unsloth\/pull\/13527<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13443\">https:\/\/github.com\/unslothai\/unsloth\/pull\/13443<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13425\">https:\/\/github.com\/unslothai\/unsloth\/pull\/13425<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13419\">https:\/\/github.com\/unslothai\/unsloth\/pull\/13419<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Unsloth v0.1.907-beta is out, bringing up to 10x faster UI responsiveness, 5x faster inference, new video models, and expanded hardware support.<\/p>\n","protected":false},"author":1,"featured_media":11445,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[1712,163,111,728,165,509,1547],"class_list":["post-11446","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-engines-and-tools","tag-flux-en","tag-gguf-en","tag-lora-en","tag-ltx-2-5-en","tag-moe-en","tag-unsloth-en","tag-verified"],"lang":"en","translations":{"en":11446,"ja":11444},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/11446","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=11446"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/11446\/revisions"}],"predecessor-version":[{"id":17757,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/11446\/revisions\/17757"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/11445"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=11446"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=11446"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=11446"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}