{"id":21563,"date":"2026-11-21T05:10:29","date_gmt":"2026-11-20T20:10:29","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/11\/21\/unsloth-v0-1-921-beta-released\/"},"modified":"2026-11-21T09:13:41","modified_gmt":"2026-11-21T00:13:41","slug":"unsloth-v0-1-921-beta-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/21\/unsloth-v0-1-921-beta-released\/","title":{"rendered":"Unsloth v0.1.921-beta Released: Faster Image Generation and MCP Chat"},"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\/v0.1.921-beta\">v0.1.921-beta<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-21<\/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&#8217;s latest version, <code>v0.1.921-beta<\/code>, has been released. Unsloth is a tool for local inference and fine-tuning of LLMs (large language models) and diffusion models (primary language: Python, license: Apache-2.0).<\/p>\n<p>The key changes in this version are faster image generation with lower memory overhead and more efficient chat processing using MCP (Model Context Protocol) tools. In particular, it includes improvements that significantly impact local inference and training users, such as better image generation performance on small GPUs and up to an 87% reduction in prompt token processing for MCP chats.<\/p>\n<h2>Key Changes<\/h2>\n<h3>Faster Image Generation and Reduced Memory Consumption<\/h3>\n<p>Processing speed and memory efficiency for image models have greatly improved. On an 8 GB small GPU environment, the new prompt rendering time for MiniMax-H3-Image has dropped from 82.7 seconds to 36.7 seconds. Additionally, system RAM usage when loading GGUF image models has been reduced by 52% to 66% on Linux and macOS.<\/p>\n<p>For model-specific optimizations, FLUX.1 has seen a 16\u201328% reduction in per-step time when streaming weights from system RAM, and FLUX.3 klein INT8 is 13% faster per step on a T4 GPU.<\/p>\n<p>Bug fixes include resolving NaN errors that occurred during image rendering on GTX 16xx cards. Furthermore, selecting a new LoRA with INT8 or FP8 models now prompts a &#8220;Reload and generate&#8221; option instead of throwing an error.<\/p>\n<h3>Reduced Token Usage and Enhanced Chat Features in MCP Chat<\/h3>\n<p>In chats utilizing MCP tools, the tool list is now fixed, improving the mechanism so that instead of reprocessing the entire conversation with every message sent, only new inputs are processed. This reduces prompt token processing volume by 83% to 87% in test environments.<\/p>\n<p>Additionally, the &#8220;Tool output length&#8221; setting under <strong>Settings &gt; Chat<\/strong> now applies to MCP results across llama.cpp, vLLM, Ollama, and Custom connections. Feature additions and improvements include OCR support for scanned PDFs within chats, preservation of paragraph shifts, date fields, and furigana text when attaching Word files, and image retention in chats imported from Claude Code or Codex. For reasoning models that hit their token limit during thought processes, advice displays and a &#8220;Retry&#8221; button have been added.<\/p>\n<h3>Video Style Transfer on Mac and Faster Video Processing<\/h3>\n<p>On Apple Silicon Mac environments, a &#8220;Restyle&#8221; feature has been added to re-render entire video clips into new styles. GPU memory required for operation is reported to be about 42 GiB for bf16 and about 31 GiB for q8_0. Execution speed for Cosmos3-Edge-FP8 has also improved by approximately 1.45x in L4 GPU environments.<\/p>\n<h3>Faster Diffusion Model LoRA and Model Training with Fixes<\/h3>\n<p>For diffusion model LoRA training, per-step time for SDXL LoRA training was reduced by 28\u201331% in T4 and L4 environments. In L4 environments, using <strong>Auto<\/strong> precision reduces FLUX.2-klein 9B training time by about 30% per step. When re-running MiniMax-H3, pre-encoded datasets are reused.<\/p>\n<p>Numerous enhancements have also been made to other training processes:<br \/>\n&#8211; AsyncGRPO step speed increased by 1.46x to 2.47x in A100 environments with TRL 1.17 and later<br \/>\n&#8211; On Strix Halo in AMD ROCm environments, sequences are now packed by default during Qwen3.5 training, increasing token processing throughput per second by about 1.4x<br \/>\n&#8211; DoRA training is accelerated by about 3\u20135% per step<\/p>\n<p>Bug fixes include resolving Gemma-4 LoRA training failures on mixed-length batches with PyTorch older than 2.14, adding PEFT 0.22 support and fixing DPO crashes when using PEFT models, and fixing an issue where input dates in JSON\/JSONL datasets unintentionally received &#8220;00:00:00&#8221;.<\/p>\n<h3>Desktop Environment and Theme Improvements<\/h3>\n<p>Theme files can now be exported and imported via <strong>Settings &gt; Appearance &gt; Theme file<\/strong>, allowing appearance settings to be shared. Bug fixes include resolving an issue where Unsloth Desktop installation would interrupt on some Windows AMD PCs, a bug where small models were incorrectly displayed as &#8220;Does not fit&#8221; in Mac chat-only GGUF environments, and an issue where users were signed out of all tabs when responses were lost during session updates.<\/p>\n<h2>Supported Models and Hardware<\/h2>\n<p>This version brings optimizations and operational support for Apple Silicon Macs, various NVIDIA and AMD GPUs\/accelerators, and diverse open-weight models. Below is an overview of what processing is now possible by hardware and model type.<\/p>\n<h3>Apple Silicon (Mac)<\/h3>\n<p>In Mac environments, the &#8220;Video Restyle&#8221; feature is now available to re-render entire video clips into new styles. Estimated GPU memory required to run this feature is about 42 GiB for bf16 precision and about 31 GiB for the q8_0 quantized version. Also, the bug where small models were incorrectly displayed as &#8220;Does not fit&#8221; in chat-only GGUF environments has been resolved.<\/p>\n<h3>NVIDIA GPU Environment<\/h3>\n<p>Inference and training performance and stability have been improved across various NVIDIA GPUs:<br \/>\n&#8211; <strong>A100<\/strong>: AsyncGRPO step speed increases by 1.46x to 2.47x when combined with TRL 1.17 or later.<br \/>\n&#8211; <strong>L4<\/strong>: Execution speed of Cosmos3-Edge-FP8 improves by about 1.45x, FLUX.2-klein 9B training speed with Auto precision improves by about 30%, and SDXL LoRA training speed is reduced by 28\u201331%.<br \/>\n&#8211; <strong>T4<\/strong>: FLUX.3 klein INT8 per-step time is reduced by 13%, and SDXL LoRA training is accelerated by 28\u201331%.<br \/>\n&#8211; <strong>GTX 16xx Series<\/strong>: NaN errors occurring during image rendering have been fixed, enabling normal image rendering.<br \/>\n&#8211; <strong>8 GB VRAM Environment<\/strong>: New prompt rendering time with MiniMax-H3-Image drops significantly from 82.7 seconds to 36.7 seconds, improving usability on lower-capacity graphics cards.<\/p>\n<h3>AMD GPU \/ ROCm Environment<\/h3>\n<p>In AMD ROCm environments utilizing processors such as Strix Halo, sequence packing is now applied by default during Qwen3.5 training. This increases processed tokens per second by about 1.4x. Additionally, an issue where installation processing failed to complete on certain Windows AMD PCs has been fixed.<\/p>\n<h3>Supported Models and Quantization Formats<\/h3>\n<p>Main model architectures and quantization formats with enhanced support and fixes in this release include:<br \/>\n&#8211; <strong>Image and Video Generation Models<\/strong>: MiniMax-H3-Image, FLUX.1 (16\u201328% faster when streaming weights from system RAM), FLUX.3 klein INT8, FLUX.2-klein 9B, Cosmos3-Edge-FP8, SDXL<br \/>\n&#8211; <strong>Language and Reasoning Models<\/strong>: Qwen3.5, Gemma-4 (batch training error fix for PyTorch &lt; 2.14), MiniMax-H3<br \/>\n&#8211; <strong>Quantization Formats<\/strong>: System RAM consumption for GGUF image models has been reduced by 52\u201366% on Linux\/macOS, and LoRA reloading and generation for INT8 and FP8 models has been streamlined.<\/p>\n<h2>How to Get It<\/h2>\n<p>Unsloth can be updated according to your execution format and OS environment. The local update procedure from the development version (<code>main<\/code> branch) and major advanced options are as follows.<\/p>\n<h3>Update Procedure for macOS, Linux, and WSL Environments<\/h3>\n<p>If you have already installed from source code, pull the latest repository, reinstall, and launch using the following commands:<\/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<p>To install and update in an isolated environment at a specific location, specify the <code>UNSLOTH_STUDIO_HOME<\/code> environment variable:<\/p>\n<pre><code class=\"language-bash\">UNSLOTH_STUDIO_HOME=&quot;$PWD\/.studio&quot;.\/install.sh --local\nUNSLOTH_STUDIO_HOME=&quot;$PWD\/.studio&quot; unsloth studio -p 8888\n<\/code><\/pre>\n<h3>Update Procedure for Windows (PowerShell) Environments<\/h3>\n<p>In PowerShell environments, set the execution policy and perform the update using similar steps:<\/p>\n<pre><code class=\"language-powershell\">cd unsloth; git pull\n.\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<p>To update in an isolated environment, configure it as follows:<\/p>\n<pre><code class=\"language-powershell\">$env:UNSLOTH_STUDIO_HOME=&quot;$PWD\\.studio&quot;;.\\install.ps1 --local\n$env:UNSLOTH_STUDIO_HOME=&quot;$PWD\\.studio&quot;; unsloth studio -p 8888\n<\/code><\/pre>\n<h3>Advanced Installer Configuration Options<\/h3>\n<p>Depending on your use case, installation with specific environment variables or flags is supported:<\/p>\n<ul>\n<li>\n<p><strong>To operate exclusively with GGUF chat without PyTorch<\/strong>:<br \/>\n  <code>bash curl -fsSL https:\/\/unsloth.ai\/install.sh | UNSLOTH_NO_TORCH=1 sh<\/code> For PowerShell: <code>powershell $env:UNSLOTH_NO_TORCH=1; irm https:\/\/unsloth.ai\/install.ps1 | iex<\/code><\/p>\n<\/li>\n<li>\n<p><strong>To specify the llama.cpp backend<\/strong> (choose from <code>vulkan<\/code>, <code>cpu<\/code>, <code>cuda<\/code>, <code>rocm<\/code>, <code>auto<\/code>):<br \/>\n  <code>bash export UNSLOTH_LLAMA_CPP_BACKEND=vulkan curl -fsSL https:\/\/unsloth.ai\/install.sh | sh<\/code><\/p>\n<\/li>\n<li>\n<p><strong>To save disk space by disabling rollback retention during installation<\/strong>:<br \/>\n  <code>bash .\/install.sh --local --no-rollback<\/code><\/p>\n<\/li>\n<\/ul>\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-v0-1-907-beta-released\/\">Unsloth v0.1.907-beta Released with Major Speed and UI Upgrades<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/20\/unsloth-v0-1-920-beta-released\/\">Unsloth v0.1.920-beta Released with Enhanced AMD NPU Support<\/a><\/li>\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<\/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 (93 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\/v0.1.921-beta\">unslothai\/unsloth Bug Fixes<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Unsloth v0.1.921-beta is out, bringing faster image generation, reduced memory usage, MCP chat token optimization, and hardware improvements.<\/p>\n","protected":false},"author":1,"featured_media":21562,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[1712,163,111,730,509,1547],"class_list":["post-21563","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-minimax-h3-en","tag-unsloth-en","tag-verified"],"lang":"en","translations":{"en":21563,"ja":21561},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21563","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=21563"}],"version-history":[{"count":1,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21563\/revisions"}],"predecessor-version":[{"id":21755,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/21563\/revisions\/21755"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/21562"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=21563"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=21563"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=21563"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}