{"id":13164,"date":"2026-11-17T11:02:12","date_gmt":"2026-11-17T02:02:12","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=13164"},"modified":"2026-11-17T11:11:20","modified_gmt":"2026-11-17T02:11:20","slug":"unsloth-v0-1-910-beta-release-2","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-910-beta-release-2\/","title":{"rendered":"Unsloth v0.1.910-beta Released with Mac MoE 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\/v0.1.910-beta\">v0.1.910-beta<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-24<\/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 Unsloth, &#8220;v0.1.910-beta&#8221;, has been released. Unsloth is a tool and UI environment for running inference and training of LLMs and diffusion models locally.<\/p>\n<p>The most significant change for users in this version is enhanced memory management for MLX MoE models on Mac environments. It is now possible to run large MoE model experts that do not fit entirely in main memory by placing them on the SSD, allowing users to run models like Qwen3.8-Flash-Next locally even on 32 GB Macs.<\/p>\n<h2>Key Changes<\/h2>\n<h3>Local Execution of Large MoE Models on Mac<\/h3>\n<p>In the MLX environment for Mac, when handling MoE (Mixture of Experts) models that exceed memory capacity, a mechanism has been introduced to keep overflowing experts on the SSD during operation. Although response speed decreases, large models like DeepSeek-V4-Flash and GLM-5.3-Flash that previously failed to start due to insufficient memory can now be executed with less memory. This impacts users who want to run large MoE models on 32 GB Macs.<\/p>\n<h3>MoE Expert Pruning Function via REAP<\/h3>\n<p>A feature has been added to the Export page for Windows and Linux to reduce and export models by removing 25% or 50% of the lowest-contributing experts using REAP (MoE Expert Pruning Method). In Python environments, using <code>reap_prune<\/code> allows processing up to 54x faster compared to llm-compressor. This is useful when you want to reduce model size and VRAM consumption in exchange for a slight decrease in model quality.<\/p>\n<h3>Lightweight Chat-only Feature and OpenAI-Compatible Fine-Tuning API<\/h3>\n<p>A &#8220;Chat-only&#8221; option has been added to the initial launch of Unsloth Desktop, installing only the GGUF-based chat functionality. By excluding training features from the initial installation, disk space usage can be significantly reduced (training features can be added later via Settings). Additionally, OpenAI-compatible fine-tuning APIs are now supported, allowing dataset uploads and SFT (Supervised Fine-Tuning) execution for local models using the OpenAI SDK.<\/p>\n<h3>Enhanced Image and Video Model Generation Features and Multi-GPU Support<\/h3>\n<p>In addition to the addition of the image model &#8220;MOSS-Image&#8221; (5.1B), a 1-step super-resolution feature &#8220;Hires fix&#8221; has been implemented for SDXL, FLUX.1, Qwen-Image, and Z-Image. Furthermore, a &#8220;Compare grid&#8221; feature to compare results side-by-side by varying seed values, CFG, and LoRA weights, as well as an AUTOMATIC1111-style prompt weighting function, are now available. Image and video models too large to fit on a single GPU are automatically split and distributed across multiple NVIDIA GPUs based on VRAM capacity.<\/p>\n<h3>Improved Training Performance and Enhanced Security<\/h3>\n<p>On NVIDIA Ampere generation and later GPUs, training processes such as DPO with array padding are now up to 3.8x faster while suppressing memory consumption. Additionally, memory usage during the training of audio models like Parakeet and Canary has been reduced by up to 55%. As a security measure, changes were made to reject revoked certificates from eMudhra during installation or updates, which could potentially be exploited by attackers for impersonation.<\/p>\n<h2>Supported Models and Hardware<\/h2>\n<p>The support status for newly supported model architectures and hardware in this version is as follows:<\/p>\n<h3>40B Coding Agent MoE Model &#8220;Flint&#8221;<\/h3>\n<p>Local execution is now supported for &#8220;Flint&#8221;, a 40B parameter (2.5B active parameter) MoE model for coding agents developed by Recursive. It supports a 256K context, reasoning features, and tool calling, and is provided in GGUF format. By selecting the 3-bit quantized version, it can run on 24 GB of VRAM during short context execution. On the SWE-bench Pro v2 benchmark, Flint scored 89.7 (compared to 73.8 for Qwen 3.6).<\/p>\n<h3>Image Generation Model Supporting English and Chinese Text Rendering &#8220;MOSS-Image&#8221;<\/h3>\n<p>Inference is now supported for &#8220;MOSS-Image&#8221; (SFT and Base versions), a 5.1B parameter image generation model developed by OpenMOSS. It is capable of generating images containing English and Chinese text. In addition to the FP8 version, a GGUF format is also available, allowing it to run on GPUs with 16 GB of VRAM as well as PC environments without a compatible GPU by using the GGUF format.<\/p>\n<h3>Tool-Calling Enhanced Model &#8220;NVIDIA Nemotron 3.5 Bolt E4B&#8221;<\/h3>\n<p>Execution and fine-tuning are now supported for &#8220;NVIDIA Nemotron 3.5 Bolt E4B&#8221;, a fine-tuned model of Gemma 4 E4B by NVIDIA. Tool-calling and instruction-following performance have been enhanced, achieving a score of 63.2 on \u03c4-bench (Tau2-Bench) compared to 38.7 for the original. Additionally, crashes that occurred during DPO execution on Gemma 4 E2B and E4B (including Nemotron 4B) have been fixed.<\/p>\n<h3>Improved NPU and GPU Accelerator Support<\/h3>\n<p>For hardware support, environments equipped with NPUs of the Ryzen AI 300 series and later have been improved to allow running models directly from the model picker without setting environment variables. Furthermore, a fix was implemented so that GGUF models, which were previously offloaded and run on the CPU in some Linux AMD PC environments with system-wide ROCm installations, now run on the GPU again.<\/p>\n<h2>How to Get It<\/h2>\n<p>Install and update using the following steps according to your Unsloth environment and use case.<\/p>\n<h3>Local Installation from Developer Version \/ Latest Source<\/h3>\n<p>To obtain and build the latest source from the repository&#8217;s <code>main<\/code> branch, run the following commands according to your OS.<\/p>\n<p><strong>For macOS \/ Linux \/ WSL:<\/strong><\/p>\n<pre><code class=\"language-bash\">git clone https:\/\/github.com\/unslothai\/unsloth\ncd unsloth\n.\/install.sh --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<p>To update an already cloned environment, update the repository 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<p>To build in an independent environment, set the environment variable <code>UNSLOTH_STUDIO_HOME<\/code>.<\/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<p><strong>For Windows (PowerShell):<\/strong><\/p>\n<pre><code class=\"language-powershell\">git clone https:\/\/github.com\/unslothai\/unsloth.git\ncd unsloth\nSet-ExecutionPolicy -Scope Process -ExecutionPolicy RemoteSigned\n.\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<p>The update procedure is as follows:<\/p>\n<pre><code class=\"language-powershell\">cd unsloth; git pull\n.\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<h3>Specifying Installation Options<\/h3>\n<p>Depending on your use case, you can modify behavior using environment variables or flags.<\/p>\n<ul>\n<li>\n<p><strong>GGUF-only mode without PyTorch:<\/strong><br \/>\n  If you want to skip the installation of PyTorch, specify <code>UNSLOTH_NO_TORCH=1<\/code>. <code>bash curl -fsSL https:\/\/unsloth.ai\/install.sh | UNSLOTH_NO_TORCH=1 sh<\/code><\/p>\n<\/li>\n<li>\n<p><strong>Skipping automatic startup:<\/strong><br \/>\n  Use <code>UNSLOTH_SKIP_AUTOSTART=1<\/code> to prevent automatic startup after setup in automation scripts, etc.<\/p>\n<\/li>\n<li>\n<p><strong>Explicitly specifying the backend:<\/strong><br \/>\n  To build with a specific backend like Vulkan, set <code>UNSLOTH_LLAMA_CPP_BACKEND<\/code>. <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>Updating the desktop app:<\/strong><br \/>\n  If you use Unsloth Desktop, obtain the latest installer corresponding to your OS from the official download page (Windows <code>.exe<\/code>, macOS <code>.dmg<\/code>, Linux <code>.deb<\/code> \/ <code>.AppImage<\/code>) and update.<\/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-v0-1-918-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<li><strong>How to read SWE-bench, \u03c4-bench<\/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\/v0.1.910-beta\">https:\/\/github.com\/unslothai\/unsloth\/releases\/tag\/v0.1.910-beta<\/a><\/li>\n<li><a href=\"https:\/\/unsloth.ai\/docs\/models\/flint\">https:\/\/unsloth.ai\/docs\/models\/flint<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/unsloth\/MOSS-Image-SFT-GGUF\">https:\/\/huggingface.co\/unsloth\/MOSS-Image-SFT-GGUF<\/a><\/li>\n<li><a href=\"https:\/\/unsloth.ai\/docs\/models\/nemotron-3.5-bolt\">https:\/\/unsloth.ai\/docs\/models\/nemotron-3.5-bolt<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/14317\">https:\/\/github.com\/unslothai\/unsloth\/pull\/14317<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/14245\">https:\/\/github.com\/unslothai\/unsloth\/pull\/14245<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Unsloth v0.1.910-beta brings Mac MLX MoE model memory management, REAP pruning, Flint and MOSS-Image support, and various updates.<\/p>\n","protected":false},"author":1,"featured_media":13163,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[3976,163,977,165,3720,509,1547],"class_list":["post-13164","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-engines-and-tools","tag-flint-en","tag-gguf-en","tag-mlx-en","tag-moe-en","tag-moss-image-en","tag-unsloth-en","tag-verified"],"lang":"en","translations":{"en":13164,"ja":13162},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13164","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=13164"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13164\/revisions"}],"predecessor-version":[{"id":17925,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13164\/revisions\/17925"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/13163"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=13164"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=13164"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=13164"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}