{"id":13593,"date":"2026-11-17T11:02:28","date_gmt":"2026-11-17T02:02:28","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=13593"},"modified":"2026-11-17T11:11:37","modified_gmt":"2026-11-17T02:11:37","slug":"unsloth-v0-1-911-beta-released-2","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/unsloth-v0-1-911-beta-released-2\/","title":{"rendered":"Unsloth v0.1.911-beta Released with Scheduled Tasks and More"},"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\/v0.1.911-beta\">v0.1.911-beta<\/a><\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-10-26<\/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 v0.1.911-beta has been released. unslothai\/unsloth is a local UI tool for running and training LLMs (Large Language Models) and diffusion models on local hardware. Developed in Python, it is available as open source under the Apache-2.0 license.<\/p>\n<p>The most significant change in this update is the addition of the &#8220;Scheduled tasks&#8221; feature for prompts. This makes it possible to automate prompt execution using specific models on a daily or weekly schedule. This is a very valuable update for users looking to leverage local models for periodic information gathering and batch processing.<\/p>\n<h2>Breaking Changes and Deprecations<\/h2>\n<p>This update includes specification changes aimed at enhancing security regarding API usage and network settings. Since existing configurations may stop working, check the following measures:<\/p>\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 style=\"text-align: left;\">Item<\/th>\n<th style=\"text-align: left;\">Old Specification<\/th>\n<th style=\"text-align: left;\">New Specification<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td style=\"text-align: left;\">Code execution approval<\/td>\n<td style=\"text-align: left;\">Self-approval allowed for Python\/terminal execution when using API keys<\/td>\n<td style=\"text-align: left;\">Self-approval not allowed for execution outside the sandbox<\/td>\n<\/tr>\n<tr>\n<td style=\"text-align: left;\">IP trust via proxy<\/td>\n<td style=\"text-align: left;\">Trusted forwarded client IPs by default<\/td>\n<td style=\"text-align: left;\">Not trusted by default (configuration required)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>If you have been having agents execute code via API, operations outside the sandbox will now be restricted. In addition, if you operate Unsloth behind a custom reverse proxy and want rate limiting to function properly, you must explicitly set the environment variable <code>UNSLOTH_STUDIO_TRUST_FORWARDED=1<\/code>. Without this setting, all users will share a single login rate limit.<\/p>\n<h2>Key Changes<\/h2>\n<h3>Introduction of Scheduled Tasks<\/h3>\n<p>You can now configure scheduled prompt execution from the &#8220;More&#8221; menu in the sidebar. Execution frequencies can be selected from once, daily, weekdays, weekly, or monthly, and execution results are automatically saved as new individual chats. These tasks can be executed using not only currently loaded models, but also other models on the device or models from connected API providers. Settings are also provided to automatically switch models if they are not loaded, as well as notification settings based on execution results (on success, on failure, or always). Note that tool usage and web searches do not function while tasks are running.<\/p>\n<h3>Support for MiniMax-M3.1-Flash<\/h3>\n<p>Inference is now supported for MiniMax&#8217;s 428B MoE (23B active parameters) model, &#8220;MiniMax-M3.1-Flash&#8221;. This model specializes in coding, tool use, and agentic workflows, supporting an extremely long context of up to 1 million tokens. In Unsloth, it is available in GGUF format, but even with the 1-bit quantized version, operation requires a combined RAM and VRAM capacity of at least 130 GB.<\/p>\n<h3>Performance Improvements on Mac (Apple Silicon)<\/h3>\n<p>MLX engine performance on Apple Silicon Macs has been significantly improved. Response speeds for MLX chats using draft models are up to 16% faster. Additionally, for MLX models with built-in draft models, the draft model is now applied automatically without requiring a separate download. For image generation, Z-Image GGUF now uses the MLX engine by default, achieving a 1.5-2x speedup compared to previous versions. Furthermore, memory release upon completion of image and video rendering has been optimized, making swaps less likely to occur on Macs with lower memory capacity.<\/p>\n<h3>Expansion and Acceleration of Image and Video Generation<\/h3>\n<p>A &#8220;Detailer&#8221; feature has been added to the image generation page. By combining SAM 3 and inpainting-compatible models, this feature redraws specific parts of generated images, such as faces and hands, at high resolution. Photo editing features using FLUX.3 Klein have also been added, enabling prompt-based editing while maintaining the original photo&#8217;s resolution. In video generation, NVIDIA&#8217;s Cosmos3-Edge is now supported, and text-to-video generation speed for Wan2.1 1.3B is about 5x faster on older NVIDIA GPUs prior to the RTX 30 series.<\/p>\n<h3>Enhancements to Training Features (LoRA\/QLoRA)<\/h3>\n<p>Features to enhance training efficiency in multi-GPU environments have been added. A &#8220;Tensor Parallelism&#8221;-like approach that splits and places each layer of a model across multiple GPUs has been introduced. By launching with <code>torchrun<\/code> and setting the number of GPUs in <code>UNSLOTH_TP_SIZE<\/code>, you can perform LoRA or QLORA training while keeping memory load per GPU low. Optimization of GRPO (Group Relative Policy Optimization) has also progressed, speeding up training steps by up to 4.4x for Llama and Qwen3 in NVIDIA GPU environments, and up to 1.9x when used in combination with vLLM.<\/p>\n<h3>Minor UI and UX Improvements<\/h3>\n<p>Execution settings can now be saved as named &#8220;presets&#8221; for each model. The next time that model is loaded (including via API), the last-used preset will be automatically applied. Furthermore, during chats using GGUF models, the progress and remaining time are now displayed while processing long prompts, removing uncertainty about wait times before responses begin. In the settings library page, storage usage is now displayed as a color-coded bar for models, cache, and free space, making management easier.<\/p>\n<h2>Hardware Requirements<\/h2>\n<p>This update brings numerous improvements to support for modern model architectures and execution efficiency on existing hardware.<\/p>\n<h3>New Supported Models and Quantization Formats<\/h3>\n<p>The most notable addition is support for MiniMax&#8217;s 428B MoE (23B active parameters) model, &#8220;MiniMax-M3.1-Flash&#8221;. This model is optimized for coding, tool use, and agentic workflows, supporting an extremely long context window of up to 1M tokens. Unsloth now supports GGUF format inference for this model. However, please note that even with the 1-bit quantized version, execution requires at least 130 GB of combined RAM and VRAM.<\/p>\n<p>For image and video generation models, text- or image-to-video generation is now supported via NVIDIA&#8217;s &#8220;Cosmos3-Edge&#8221;. For the video generation model &#8220;Wan2.1 1.3B&#8221;, rendering speed has been improved by about 5x on older NVIDIA GPUs prior to the RTX 30 series. Image editing now supports &#8220;FLUX.3 Klein&#8221;, enabling prompt-based editing (inpaint) while maintaining original photo resolutions.<\/p>\n<h3>Optimization for Apple Silicon (Mac)<\/h3>\n<p>For Mac users, this update brings significant benefits. Text-to-image generation using Z-Image GGUF now uses the MLX engine by default, achieving a 1.5-2x speedup over previous performance. Furthermore, response speeds for MLX chats using draft models have increased by up to 16%. MLX models with built-in draft models automatically utilize this feature without additional downloads. Additionally, environments using torch 2.13 or later now invoke the Apple MPS (Metal Performance Shaders) SDPA prefill kernel on a per-call basis, further streamlining processing efficiency.<\/p>\n<h3>Multi-GPU and Training Optimizations<\/h3>\n<p>For engineers conducting large-scale training, LoRA\/QLoRA training features in multi-GPU environments have been enhanced. By setting the number of GPUs in the environment variable <code>UNSLOTH_TP_SIZE<\/code> and launching with <code>torchrun<\/code>, model layers can be split across GPUs, reducing VRAM load per card. Native FSDP2 (Fully Sharded Data Parallel) now supports sharded resume loading for weights and optimizer states, making it more efficient to resume training for large models. Context parallelism (Ulysses) has also been introduced for DPO and GRPO training, enabling training on long sequences.<\/p>\n<h3>Other Hardware Support and Fixes<\/h3>\n<p>An issue in Windows environments where loading GGUF models failed with the error &#8220;llama-server failed to start&#8221; has been fixed. Additionally, a fallback function has been implemented to start training on CPU instead of rejecting imports when no GPU is detected or in non-GPU environments. On NVIDIA GPUs, routing to factored ConvRot kernels has been optimized for specific architectures such as sm86, sm90, sm100, and sm121, improving image generation latency.<\/p>\n<h2>How to Get It<\/h2>\n<p>Please update according to your usage format using the steps below.<\/p>\n<h3>Unsloth Desktop (GUI App)<\/h3>\n<p>Download and run the latest installer for your OS from the links below:<\/p>\n<ul>\n<li><strong>Windows<\/strong>: <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-Windows.exe\">Download<\/a><\/li>\n<li><strong>macOS<\/strong>: <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-MacOS.dmg\">Download<\/a><\/li>\n<li><strong>Linux (Ubuntu deb)<\/strong>: <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-Ubuntu.deb\">Download<\/a><\/li>\n<li><strong>Linux (AppImage)<\/strong>: <a href=\"https:\/\/github.com\/unslothai\/unsloth\/releases\/latest\/download\/Unsloth-Desktop-Linux.AppImage\">Download<\/a><\/li>\n<\/ul>\n<h3>Unsloth Studio (Developer\/Nightly Edition)<\/h3>\n<p>If you are using a cloned repository, fetch the latest source code and run the installation script with the following commands:<\/p>\n<p><strong>For macOS, Linux, and WSL:<\/strong><\/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><strong>For Windows (PowerShell):<\/strong><\/p>\n<pre><code class=\"language-powershell\">cd unsloth; git pull\n.\\install.ps1 --local\nunsloth studio -p 8888\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-v0-1-915-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<\/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\/v0.1.911-beta\">https:\/\/github.com\/unslothai\/unsloth\/releases\/tag\/v0.1.911-beta<\/a><\/li>\n<li><a href=\"https:\/\/unsloth.ai\/docs\/models\/minimax-m3.1-flash\">https:\/\/unsloth.ai\/docs\/models\/minimax-m3.1-flash<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/unsloth\/MiniMax-M3.1-Flash-GGUF\">https:\/\/huggingface.co\/unsloth\/MiniMax-M3.1-Flash-GGUF<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/14940\">https:\/\/github.com\/unslothai\/unsloth\/pull\/14940<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/14941\">https:\/\/github.com\/unslothai\/unsloth\/pull\/14941<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/14616\">https:\/\/github.com\/unslothai\/unsloth\/pull\/14616<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/15042\">https:\/\/github.com\/unslothai\/unsloth\/pull\/15042<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/15041\">https:\/\/github.com\/unslothai\/unsloth\/pull\/15041<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/unslothai\/unsloth\/pull\/14717\">https:\/\/github.com\/unslothai\/unsloth\/pull\/14717<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Unsloth v0.1.911-beta is out, introducing scheduled tasks, MiniMax-M3.1-Flash support, Apple Silicon optimizations, and multi-GPU training improvements.<\/p>\n","protected":false},"author":1,"featured_media":13592,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[505,163,111,4128,977,509,1547],"class_list":["post-13593","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-engines-and-tools","tag-apple-silicon-en","tag-gguf-en","tag-lora-en","tag-minimax-m3-1-flash-en","tag-mlx-en","tag-unsloth-en","tag-verified"],"lang":"en","translations":{"en":13593,"ja":13591},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13593","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=13593"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13593\/revisions"}],"predecessor-version":[{"id":17944,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/13593\/revisions\/17944"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/13592"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=13593"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=13593"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=13593"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}