{"id":11954,"date":"2026-11-17T11:00:50","date_gmt":"2026-11-17T02:00:50","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=11954"},"modified":"2026-11-17T11:09:58","modified_gmt":"2026-11-17T02:09:58","slug":"unsloth-v0-1-908-beta-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-908-beta-released\/","title":{"rendered":"Unsloth v0.1.908-beta Released with Nemotron 3.5 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\/v0.1.908-beta\">v0.1.908-beta<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-16<\/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 v0.1.908-beta has been released. Unsloth is a UI tool for running and training LLMs (Large Language Models) and diffusion models in local environments.<\/p>\n<p>The most significant updates in this release include support for NVIDIA&#8217;s latest &#8220;Nemotron 3.5&#8221; series and Meta&#8217;s &#8220;Muse Spark Open&#8221;, as well as the addition of training and chat features for text-diffusion LLMs. This greatly expands the options for handling very large MoE (Mixture of Experts) models and new diffusion model architectures locally.<\/p>\n<h2>Key Changes<\/h2>\n<h3>Support for NVIDIA Nemotron 3.5 and Meta Muse Spark Open<\/h3>\n<p>NVIDIA&#8217;s new Nemotron 3.5 models and Meta&#8217;s Muse Spark Open can now run on Unsloth.<\/p>\n<p>NVIDIA Nemotron 3.5 Super VL is a 120B (12B active parameter) model supporting up to 1M tokens of context. Memory usage is approximately 64GB for 1-bit and approximately 128GB for 4-bit. Additionally, Nemotron 3.5 Bolt (74B, 9B active parameters, 128K context), which is compressed from Super VL, runs on approximately 32GB of memory in 1-bit or 2-bit. By default, Nemotron 3.5 Bolt uses the DSpark drafter and Super VL uses built-in MTP for fast output, though these can be turned off in the settings.<\/p>\n<p>Furthermore, Meta&#8217;s Muse Spark Open is an 816B MoE model with approximately 42.7B active parameters and a 1M token context. It runs using approximately 256GB of memory in 2-bit. These models are built for multimodal reasoning, coding, and agentic workflows.<\/p>\n<h3>Training and Chat for Text-Diffusion LLMs<\/h3>\n<p>Training and chatting with text-diffusion models such as LLaDA, Dream, SDAR, LLaDA2, Nemotron-Labs-Diffusion, and DiffusionGemma are now available within the Unsloth app. A notable feature is the display of a live denoising canvas when generating chat responses.<\/p>\n<h3>Enhanced Training Features via DeepSpeed<\/h3>\n<p>Training using DeepSpeed ZeRO-2 and ZeRO-3 is now supported. This includes 4-bit QLoRA, full fine-tuning, DPO, and CPU offloading. Improvements have also been made to prevent crashes when using gradient checkpointing in LoRA training with DDP (Distributed Data Parallel).<\/p>\n<h3>Prompt Caching Improvements for Coding Agents<\/h3>\n<p>The prompt caching mechanism has been improved when using coding agents like Claude Code or Codex with local GGUF models. This significantly reduces the amount of prompt data reprocessed when continuing conversations. Specifically, it is reported to make the first response start of a new Claude Code session up to 4x faster.<\/p>\n<h3>Improved Inference Performance<\/h3>\n<p>In the Python environment, single-prompt <code>model.generate<\/code> with Llama and Qwen3 models at 32K context is up to 2.7x faster. Additionally, on certain NVIDIA GPUs, using bitsandbytes 4-bit models to generate 2 to 4 prompts simultaneously has been observed to be up to 2.8x faster.<\/p>\n<h3>Expanded Browser Use and Image Generation<\/h3>\n<p>In the &#8220;Browser use&#8221; feature, a cursor can now be displayed when the model performs clicks and types (this can be disabled in settings). A &#8220;Always allow&#8221; setting per site has also been added.<\/p>\n<p>For image generation, images can now be generated from the Images page through connected image servers such as sd-server, Ollama, Lemonade, vLLM, llama.cpp, or custom OpenAI-compatible URLs.<\/p>\n<h2>Supported Models and Hardware<\/h2>\n<p>With this update, support for the following model architectures and quantization formats has been enhanced.<\/p>\n<h3>Supported Models<\/h3>\n<ul>\n<li><strong>NVIDIA Nemotron 3.5 Series<\/strong>:\n<ul>\n<li><strong>Nemotron 3.5 Super VL<\/strong>: 120B (12B active) model. Aimed at document comprehension and computer use. Uses approximately 64GB of memory for 1-bit and approximately 128GB for 4-bit. &#8211; <strong>Nemotron 3.5 Bolt<\/strong>: 74B (9B active) model. A compressed version of Super VL, running on approximately 32GB of memory in 1-bit or 2-bit.<\/li>\n<\/ul>\n<\/li>\n<li><strong>Meta Muse Spark Open<\/strong>: 816B MoE model (42.7B active). Aimed at multimodal reasoning and coding, using approximately 256GB of memory in 2-bit.<\/li>\n<li><strong>Text-Diffusion LLMs<\/strong>: Supports training and chatting for LLaDA, Dream, SDAR, LLaDA2, Nemotron-Labs-Diffusion, DiffusionGemma, and others.<\/li>\n<li><strong>MLX Repository Quantization Formats<\/strong>: Similar to GGUF, users can select 2-bit, 4-bit, 6-bit, and 8-bit precisions published by MLX repositories.<\/li>\n<\/ul>\n<h3>Hardware and Compatibility<\/h3>\n<ul>\n<li><strong>AMD GPU<\/strong>: GPUs like RDNA 1 and Polaris, which are not supported by ROCm, can now run using the Vulkan llama.cpp build.<\/li>\n<li><strong>Windows Environment<\/strong>: When using NVIDIA drivers older than CUDA 12, the system will install Vulkan or CPU llama.cpp builds instead of failing with an error.<\/li>\n<li><strong>Multi-GPU<\/strong>: Machines mixing different generations of GPUs can now select appropriate attention backends suited for each GPU.<\/li>\n<\/ul>\n<h2>How to Update<\/h2>\n<p>If you are using the Unsloth Studio (Web UI version) developer or nightly builds, update using the following steps.<\/p>\n<h3>For macOS, Linux, and WSL<\/h3>\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<pre><code class=\"language-powershell\">cd unsloth; git pull\n.\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<p>For detailed installation steps, refer to the <a href=\"https:\/\/unsloth.ai\/docs\/new\/studio\/install#uninstall\">official documentation<\/a>.<\/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-v0-1-918-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\/v0.1.908-beta\">unslothai\/unsloth NVIDIA Nemotron 3.5 + Meta Muse Spark Open (GitHub Release)<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Unsloth v0.1.908-beta adds support for NVIDIA Nemotron 3.5, Meta Muse Spark Open, text-diffusion LLMs, and DeepSpeed training.<\/p>\n","protected":false},"author":1,"featured_media":11953,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[3495,163,165,2696,509,1547],"class_list":["post-11954","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-engines-and-tools","tag-deepspeed-en","tag-gguf-en","tag-moe-en","tag-nemotron-en","tag-unsloth-en","tag-verified"],"lang":"en","translations":{"en":11954,"ja":11952},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/11954","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=11954"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/11954\/revisions"}],"predecessor-version":[{"id":17815,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/11954\/revisions\/17815"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/11953"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=11954"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=11954"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=11954"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}