{"id":12585,"date":"2026-11-17T11:01:25","date_gmt":"2026-11-17T02:01:25","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=12585"},"modified":"2026-11-17T11:10:31","modified_gmt":"2026-11-17T02:10:31","slug":"unsloth-v0-1-909-beta-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-909-beta-released\/","title":{"rendered":"Unsloth v0.1.909-beta Released with DeepSeek-V4.1-Flash 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\/v0.1.909-beta\">v0.1.909-beta<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-20<\/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 version &#8220;v0.1.909-beta&#8221;, a tool for local execution and fine-tuning of large language models (LLMs) and diffusion models, has been released. Unsloth is Python-developed Apache-2.0 licensed software that supports efficient inference using GGUF and MLX formats, alongside fast additional training for various models.<\/p>\n<p>The main highlights of this update are official support for the local execution of the latest model &#8220;DeepSeek-V4.1-Flash-1017&#8221; and the image generation model &#8220;FLUX.3 Image Klein&#8221;. It enables memory-efficient local execution of ultra-large models and image editing using multiple reference images. Significant functional expansions have been made to both inference and training, including chat integration for image\/video generation and multi-GPU training support via FSDP2.<\/p>\n<h2>Key Changes<\/h2>\n<h3>Support for DeepSeek-V4.1-Flash-1017<\/h3>\n<p>It is now possible to run &#8220;DeepSeek-V4.1-Flash-1017&#8221; locally, which features improved inference capabilities, enhanced coding and agent performance, and more concise outputs. The total parameter count consists of 552B (approx. 16B active) plus a 196B Engram table, with support for a 1M token context length. By placing the Engram table on an SSD, it operates with approximately 125GB of memory under 1-bit quantization and about 300GB under 4-bit quantization. The score on &#8220;Terminal-Bench 4.0&#8221;, a benchmark evaluating terminal operation tasks, has improved from 42.3 compared to 31.2 for the previous DeepSeek-V4.1-Flash. The thinking function is enabled by default (high effort) and supports intensity adjustments from low to max as well as disabling.<\/p>\n<h3>Image Generation and Editing with FLUX.3 Image Klein<\/h3>\n<p>Image generation and editing are now supported using Black Forest Labs&#8217; 5.6B parameter model &#8220;FLUX.3 Image Klein&#8221;. It allows text rendering and editing using up to 10 reference images, operating on 24GB of GPU memory. The provided GGUF version also works on PCs without a compatible dedicated GPU, and the non-distilled Klein Base is available as well. The &#8220;Enhance prompt&#8221; feature, which automatically expands short prompts into detailed captions, is enabled by default to suppress body distortions that tend to occur with complex poses.<\/p>\n<h3>Simultaneous Retention and Integration of Image\/Video Models in Chat<\/h3>\n<p>In environments equipped with dedicated GPUs, it is now possible to keep image, video, and text-to-speech models resident in memory alongside chat models. You can configure this behavior by selecting &#8220;When they fit&#8221; or &#8220;Always&#8221; under &#8220;Settings &gt; Resources &gt; Model memory&#8221;. Enabling the &#8220;Images&#8221; toggle allows the local chat model to call resident image and video models to generate images and playable videos directly within the chat interface. Even if a model swap occurs, enabled &#8220;Keep chat context across unloads&#8221; lets you continue the conversation without reloading the context.<\/p>\n<h3>Addition of MCP Server Settings for Coding Agents<\/h3>\n<p>Coding agents such as Claude Code, Codex, and Cursor can now be connected to Unsloth&#8217;s MCP server with a single command. Configuration is completed by running <code>unsloth mcp add claude-code<\/code> (or <code>codex<\/code>, <code>cursor<\/code>). Connected agents can upload datasets, start training, and load generated checkpoints, as well as issue instructions to pre-download models, preview download\/unload contents upon model loading, and queue video generation jobs.<\/p>\n<h3>Multi-GPU Training via FSDP2 and Context Parallelism<\/h3>\n<p>Training via Python code now supports multi-GPU distribution for LoRA, QLoRA, and full fine-tuning using FSDP2. Additionally, by specifying <code>context_parallel_size<\/code>, long contexts containing packed sequences can be split and processed across multiple GPUs. A preflight command for distributed environments, <code>unsloth doctor --gpus<\/code>, has also been added. Furthermore, the introduction of <code>optim = \"adamw_flash\"<\/code> to reduce optimizer memory during full fine-tuning, and offloading optimizer states to main memory via the environment variable <code>UNSLOTH_OFFLOAD_OPTIMIZER_STATE=1<\/code>, are now available.<\/p>\n<h3>Inference and Training Acceleration for Tool Execution and Various Architectures<\/h3>\n<p>In local GGUF chats involving tool calls, responses are now displayed instantly at the speed the model generates tokens, eliminating tool call wait times and speed drops during long text outputs. On the training side, LoRA training using <code>lora_dropout<\/code> or <code>lora_bias<\/code> now supports Unsloth&#8217;s fused kernels, accelerating it by approximately 28% per step. Additionally, a Triton kernel running on NVIDIA Ampere and newer architectures has been added for Gemma 4&#8217;s global attention layers, speeding up attention processing by 1.4x to 2.1x while reducing memory usage by up to 4.9x. Optimizations have also been made for Apple Silicon environments (MLX), improving processing speeds for Qwen3.8-Flash-Next by up to 1.4x on M5 Macs.<\/p>\n<h3>Improved Stability and UI Enhancements in Unsloth Desktop<\/h3>\n<p>In Unsloth Desktop, you can now select existing working folders when creating a project. Furthermore, a bug causing crashes on startup in older macOS environments has been fixed, along with support for browser panel behavior on macOS 27 and adherence to system proxies like Clash.<\/p>\n<h2>Specifications<\/h2>\n<h3>Newly Supported Models and Image\/Video Processing Architectures<\/h3>\n<p>This release supports inference and editing for the latest large language models and image generation models.<\/p>\n<ul>\n<li><strong>DeepSeek-V4.1-Flash-1017<\/strong>: Local GGUF inference is now supported for this latest model, featuring a total parameter count of 552B (approx. 16B active) and a 196B Engram table. With the Engram table placed on an SSD, it can run in memory environments of about 125GB for 1-bit quantization and about 300GB for 4-bit quantization.<\/li>\n<li><strong>FLUX.3 Image Klein \/ Klein Base<\/strong>: Support added for the 5.6B parameter image generation and editing model developed by Black Forest Labs. It runs on 24GB of GPU memory, and GGUF formats allow inference even on PC environments without a supported dedicated GPU. Non-distilled Klein Base GGUF versions are also provided.<\/li>\n<li><strong>Qwen-Image-Edit-2509 \/ 2511<\/strong>: Up to 3 reference images can now be accepted simultaneously in a single editing operation.<\/li>\n<li><strong>LTX-2.5<\/strong>: The Hugging Face login authentication previously required during use is no longer necessary.<\/li>\n<li><strong>Gemma 4<\/strong>: Added a Triton kernel (<code>head_dim 512<\/code>) to fast-process global attention layers for NVIDIA Ampere and later GPU environments.<\/li>\n<\/ul>\n<h3>Optimization for Apple Silicon Environments<\/h3>\n<p>Execution efficiency on Apple Silicon Macs (MLX) has been improved. Operation speeds increase by up to 1.4x for Qwen3.8-Flash-Next on M5 chip Mac environments, and up to 1.2x on M1 through M4 Mac environments. Furthermore, simultaneous generation processing for multiple chats using 4-bit and 8-bit MLX models is up to 1.8x faster on M1 through M4 Mac environments.<\/p>\n<h3>Expanded Hardware Compatibility and Platform-Specific Adjustments<\/h3>\n<p>Execution backends have been expanded and stabilized for diverse hardware environments.<\/p>\n<ul>\n<li><strong>Vulkan Support on Snapdragon and Linux ARM Boards<\/strong>: On PCs lacking NVIDIA GPUs or ROCm environments (such as Snapdragon-powered laptops and Linux ARM boards), GGUF format models can now run on GPUs via the Vulkan backend without relying on CPU processing.<\/li>\n<li><strong>Support for NVIDIA Multi-GPU Mixed Environments and MIG<\/strong>: NVIDIA MIG (Multi-Instance GPU) slices are now correctly recognized and selected by exact size. Additionally, in machine environments where multiple GPUs of different generations are mixed, settings and llama.cpp builds that allow all installed GPUs to operate normally are applied.<\/li>\n<li><strong>Improvements for AMD Processor and GPU Environments<\/strong>: Linux machines equipped with both AMD Strix APUs and discrete AMD GPUs have been improved so that PyTorch is properly installed for the discrete GPU. Also, the issue where administrator privileges were requested upon opening the settings screen (<code>Settings &gt; System<\/code>) on Windows AMD PCs with the HIP SDK installed has been resolved.<\/li>\n<li><strong>Windows on ARM<\/strong>: &#8220;Sandbox High&#8221; can now be enabled under <code>Settings &gt; Sandbox<\/code> in the settings screen.<\/li>\n<\/ul>\n<h2>How to Get It<\/h2>\n<p>Update procedures for Unsloth vary depending on your usage form (Unsloth Desktop, Unsloth Studio, Unsloth Core).<\/p>\n<h3>Updating Unsloth Desktop<\/h3>\n<p>If you are using the desktop application, please download and apply the latest installer for your platform.<\/p>\n<ul>\n<li>Windows: <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-Windows.exe\">Unsloth-Desktop-Windows.exe<\/a><\/li>\n<li>Windows ARM64: <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/Unsloth-Desktop-Windows-ARM64.exe\">Unsloth-Desktop-Windows-ARM64.exe<\/a><\/li>\n<li>macOS: <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-MacOS.dmg\">Unsloth-Desktop-MacOS.dmg<\/a><\/li>\n<li>Linux x64 (Ubuntu deb): <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-Ubuntu.deb\">Unsloth-Desktop-Ubuntu.deb<\/a><\/li>\n<li>Linux ARM64 (Ubuntu 24.04+ deb): <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-Ubuntu-ARM64.deb\">Unsloth-Desktop-Ubuntu-ARM64.deb<\/a><\/li>\n<li>Linux x64 (AppImage): <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-Linux.AppImage\">Unsloth-Desktop-Linux.AppImage<\/a><\/li>\n<\/ul>\n<h3>Updating Source Code (Development \/ Local Environments)<\/h3>\n<p>If you have cloned the repository and are running Unsloth Studio in a local environment, fetch the latest commits from the repository and run the update script.<\/p>\n<p>Update steps for macOS, Linux, and WSL environments:<\/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>Update steps for Windows (PowerShell) environments:<\/p>\n<pre><code class=\"language-powershell\">cd unsloth; git pull\n.\\install.ps1 --local\nunsloth studio -p 8888\n<\/code><\/pre>\n<p>If you want to install or update while explicitly specifying Vulkan as the backend, you can run the script with the environment variable set.<\/p>\n<p>For macOS, Linux, and WSL environments:<\/p>\n<pre><code class=\"language-bash\">export UNSLOTH_LLAMA_CPP_BACKEND=vulkan\ncurl -fsSL https:\/\/unsloth.ai\/install.sh | sh\n<\/code><\/pre>\n<p>For Windows (PowerShell) environments:<\/p>\n<pre><code class=\"language-powershell\">$env:UNSLOTH_LLAMA_CPP_BACKEND=&quot;vulkan&quot;\nirm https:\/\/unsloth.ai\/install.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<li><strong>How to read Terminal-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\/v0.1.909-beta\">unslothai\/unsloth Release Notes (v0.1.909-beta)<\/a><\/li>\n<li><a href=\"https:\/\/unsloth.ai\/docs\/models\/deepseek-v4.1-flash\">DeepSeek-V4.1-Flash Official Guide<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/unsloth\/DeepSeek-V4.1-Flash-1017-GGUF\">DeepSeek-V4.1-Flash-1017 GGUF (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/unsloth\/flux-3-image-klein-GGUF\">FLUX.3 Image Klein GGUF (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/unsloth\/flux-3-image-klein-base-GGUF\">FLUX.3 Image Klein Base GGUF (Hugging Face)<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13865\">PR #13865: FSDP2: first-class LoRA and full finetuning<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13869\">PR #13869: Add a Triton head_dim 512 attention kernel for Gemma-4 global layers<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13881\">PR #13881: Add optim = &#8220;adamw_flash&#8221; for lower optimizer memory<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13882\">PR #13882: Add opt-in optimizer state CPU offload<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13884\">PR #13884: Add unsloth doctor &#8211;gpus multi-GPU preflight<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13898\">PR #13898: Fused LoRA kernels: support lora_dropout and lora_bias<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13928\">PR #13928: Studio: automatic image\/video loads pick the freest GPU<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/13931\">PR #13931: Installers (AMD, Linux): Strix APU defers torch to discrete card<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Unsloth v0.1.909-beta brings local support for DeepSeek-V4.1-Flash-1017 and FLUX.3 Image Klein, FSDP2 multi-GPU training, MCP server integration, and more.<\/p>\n","protected":false},"author":1,"featured_media":12584,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[3738,3740,3742,163,977,509,1547],"class_list":["post-12585","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-engines-and-tools","tag-deepseek-v4-1-en","tag-flux-3-en","tag-fsdp2-en","tag-gguf-en","tag-mlx-en","tag-unsloth-en","tag-verified"],"lang":"en","translations":{"en":12585,"ja":12583},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12585","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=12585"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12585\/revisions"}],"predecessor-version":[{"id":17861,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/12585\/revisions\/17861"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/12584"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=12585"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=12585"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=12585"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}