{"id":16472,"date":"2026-11-17T11:04:47","date_gmt":"2026-11-17T02:04:47","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=16472"},"modified":"2026-11-17T11:14:15","modified_gmt":"2026-11-17T02:14:15","slug":"unsloth-v0-1-918-beta-release","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-918-beta-release\/","title":{"rendered":"Unsloth v0.1.918-beta Adds Computer Use, MCP Game Engine Support"},"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\/v0.1.918-beta\">v0.1.918-beta<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-17<\/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>The latest version of unslothai\/unsloth, v0.1.918-beta, has been released. unslothai\/unsloth is an open-source (Apache-2.0 license) Python-based local tool for running and training LLMs and diffusion models.<\/p>\n<p>The most significant change in this version is the addition of &#8220;Computer use (beta),&#8221; a feature that allows the model to recognize your PC screen and operate applications using locally running vision models. This enables automation and interaction with various applications via local models.<\/p>\n<h2>Key Changes<\/h2>\n<h3>Introduction of Computer use (beta)<\/h3>\n<p>You can now view your screen and control your mouse and keyboard using locally running vision models (GGUF, safetensors, or MLX formats) or connected providers. It is compatible with macOS, Windows, and Linux. To use it, enable the feature in &#8220;Settings &gt; Computer use&#8221; and select it from the composer&#8217;s + menu.<\/p>\n<p>For safety, a confirmation dialog appears during the first operation, when using a new application, and before risky steps like terminal operations. Operations on blocked apps are strictly prohibited. In an emergency, you can regain control by pressing the Stop button or moving the pointer to the corner of the screen. In desktop environment testing, Qwen3.8-27B is reported to have shown the best results.<\/p>\n<h3>MCP Integration with Game Engines and Creative Tools<\/h3>\n<p>In the MCP (Model Context Protocol) menu, Unity, Unreal Engine, Godot, Figma, and Roblox are now available in addition to the existing Blender support. This allows you to build scenes and games using local models.<\/p>\n<p>To use, select the target app, check the permission box, run &#8220;Test connection,&#8221; and click &#8220;Enable.&#8221; Note that Figma access is read-only, allowing only the reference of frames, styles, and assets. Demos have shown successful construction of a Pong game in Godot and a village square in Unreal Engine using Qwen3.8-27B.<\/p>\n<h3>Support for MiniMax-H3-Image and Performance Optimizations<\/h3>\n<p>MiniMax-H3-Image models can now run in GGUF format and at FP8\/INT8 precision. The GGUF version runs on stable-diffusion.cpp, enabling execution on CPU, Mac, small GPUs, and AMD environments. FP8 and INT8 weights are provided for NVIDIA environments; FP8 requires a GPU with SM 8.9 or higher. Selecting &#8220;Auto&#8221; in Precision settings will attempt INT8, FP8, and bf16 in order, downloading only the necessary weights.<\/p>\n<p>Additionally, LoRA training and loading for MiniMax-H3-Image are now supported. In terms of performance, the 2048 px image decoding time on an A100 80GB environment has been significantly reduced from 2.82 seconds to 0.28 seconds.<\/p>\n<h3>Refreshed Model Picker and Enhanced Project Management<\/h3>\n<p>The UI on the composer has been refreshed, replacing the old &#8220;Thinking pill&#8221; with model chips that display the model&#8217;s logo, name, and reasoning level. Clicking this allows you to switch models or adjust reasoning levels from tabs such as Favorites, On Device, Discover, and Connected without leaving the chat screen.<\/p>\n<p>A new folder system has been introduced for project management. A Files tab has been added, displaying working folders and git branches for each project, allowing you to browse folder contents, search files, and check git change history.<\/p>\n<h3>Expanded Connection Providers and Auto-Compaction<\/h3>\n<p>For connections such as OpenRouter, DeepSeek, Gemini, vLLM, and Ollama, an auto-compaction feature has been implemented that summarizes past turns during long chats instead of deleting them. You can set the timing for compaction in the &#8220;Auto-compact after&#8221; field in connection settings. You can also manually execute the <code>\/compact<\/code> command.<\/p>\n<p>The connection gallery has been refreshed with live status indicators and nine new providers: xAI, NVIDIA NIM, Perplexity, Cohere, Vercel AI Gateway, Azure OpenAI, Amazon Bedrock, Google Vertex AI, and SGLang.<\/p>\n<h3>Enhanced Training and Performance Improvements<\/h3>\n<p>In training, a new attention kernel has been adopted for Gemma 4 12B, improving training speed by 1.4\u20132.6x for 2048\u20138192 tokens. DeepSeek-V4.1 now supports native LoRA and QLoRA, as well as FSDP2, tensor parallelism, and pipeline parallelism.<\/p>\n<p>For SFT (Supervised Fine-Tuning) epoch-based execution, text dataset packing is now enabled by default (set <code>packing = False<\/code> to disable). LoRA training for FLUX.3 Klein and Qwen-Image-2.1 is now supported, and new INT8 attention has improved FLUX.3 Klein (2048 px) rendering speed by 15\u201323% on RTX PRO 6000, A100, and L4 environments.<\/p>\n<h2>Supported Models and Hardware<\/h2>\n<h3>MiniMax-H3-Image<\/h3>\n<p>MiniMax-H3-Image GGUF now runs on stable-diffusion.cpp, allowing execution on CPU, Mac, small GPUs, and AMD environments. FP8 and INT8 weights are available for NVIDIA environments. FP8 requires a GPU with SM 8.9 or higher. GGUF is distributed at <a href=\"https:\/\/huggingface.co\/unsloth\/MiniMax-H3-Image-GGUF\">MiniMax-H3-Image-GGUF<\/a>, and FP8\/INT8 at <a href=\"https:\/\/huggingface.co\/unsloth\/MiniMax-H3-Image-FP8\">MiniMax-H3-Image-FP8<\/a>.<\/p>\n<h3>Other Models<\/h3>\n<ul>\n<li>Inkling-2-Small can be run as <a href=\"https:\/\/huggingface.co\/unsloth\/Inkling-2-Small-GGUF\">Dynamic GGUFs<\/a>.<\/li>\n<li>DeepSeek-V4.1 can be loaded via transformers port and trained with LoRA \/ QLoRA.<\/li>\n<li>LoRA for FLUX.3 Klein and Qwen-Image-2.1 can be trained. Qwen-Image-2.1 LoRA can be run on Qwen-Image-2.1-Turbo.<\/li>\n<li>Qwen3.8-Flash-Next has a fix where layers remain compiled during block compilation.<\/li>\n<\/ul>\n<h3>Hardware<\/h3>\n<ul>\n<li>INT8 attention and CUTLASS int8 GEMM with fused dequant epilogue have been added. FLUX.3 Klein 2048 px rendering is reported to be 15\u201323% faster on RTX PRO 6000, A100, and L4.<\/li>\n<li>Vulkan opt-in for sd.cpp is now reflected for Intel GPUs on Windows.<\/li>\n<li>A fix is included for an issue where the CUDA bundle did not include cudart\/cublas in chat-only installations, causing llama.cpp to run on the CPU (PR #17463). If you are in this environment, please verify that the GPU is being used after updating.<\/li>\n<\/ul>\n<h2>How to Update<\/h2>\n<p>The update procedure for Unsloth Studio as described in the README is as follows. This is a developer installation built from the <code>main<\/code> branch (nightly), and <code>git pull<\/code> may result in a state newer than the release tag.<\/p>\n<p>macOS \/ Linux \/ WSL:<\/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>Windows PowerShell:<\/p>\n<pre><code class=\"language-powershell\">cd unsloth; git pull\n.\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<p>For fresh installations or to use Unsloth Desktop, you can download it from <a href=\"https:\/\/unsloth.ai\/download\">unsloth.ai\/download<\/a>.<\/p>\n<p>If you only use GGUF, you can choose a mode without PyTorch:<\/p>\n<pre><code class=\"language-bash\">curl -fsSL https:\/\/unsloth.ai\/install.sh | UNSLOTH_NO_TORCH=1 sh\n<\/code><\/pre>\n<p>During reinstallation, the old and new environments coexist temporarily, requiring space for two installations. If you lack space, adding <code>--no-rollback<\/code> requires only enough space for one, but you cannot revert if the installation fails.<\/p>\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-v0-1-915-beta-released\/\">Unsloth v0.1.915-beta Released with Browser Use &amp; Qwen3.8<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-913-beta-release\/\">Unsloth v0.1.913-beta Released: Faster Updates and Image Gen<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-0-1-912-beta-updates\/\">Unsloth 0.1.912-beta: Multi-LoRA Batching and Faster Startup<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-911-beta-released-2\/\">Unsloth v0.1.911-beta Released with Scheduled Tasks and More<\/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<li><strong>How to read Perplexity<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-benchmarks-en\/\">Benchmark glossary<\/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\/v0.1.918-beta\">unslothai\/unsloth v0.1.918-beta release page<\/a><\/li>\n<li><a href=\"https:\/\/unsloth.ai\/docs\/basics\/computer-use\">Computer use guide<\/a><\/li>\n<li><a href=\"https:\/\/unsloth.ai\/docs\/basics\/mcp\/mcp-engines\">Game engine MCP guide<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/unsloth\/MiniMax-H3-Image-GGUF\">MiniMax-H3-Image-GGUF<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/unsloth\/MiniMax-H3-Image-FP8\">MiniMax-H3-Image-FP8<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/unsloth\/Inkling-2-Small-GGUF\">Inkling-2-Small-GGUF<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/17463\">PR #17463<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Unsloth v0.1.918-beta introduces local Computer use, expanded MCP support for game engines, MiniMax-H3-Image optimizations, and improved training speed.<\/p>\n","protected":false},"author":1,"featured_media":16471,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[5144,1133,163,111,4846,509,1547],"class_list":["post-16472","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-engines-and-tools","tag-computer-use-en","tag-fp8-en","tag-gguf-en","tag-lora-en","tag-minimax-h3-image-en","tag-unsloth-en","tag-verified"],"lang":"en","translations":{"en":16472,"ja":16470},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/16472","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=16472"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/16472\/revisions"}],"predecessor-version":[{"id":18135,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/16472\/revisions\/18135"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/16471"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=16472"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=16472"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=16472"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}