Unsloth

What this means for you: Unsloth, the app for running AI models on your own computer, now lets you change how it looks. You can pick one of five Glass themes, set a custom greeting (text, an icon or an image), adjust the background, and choose code fonts with line numbers and word wrapping. On a Mac, the Desktop app can let your desktop show through the window.

Why it matters: It follows the Oct 30 update that added custom chat backgrounds and a movable message box, so this continues the same run of look-and-feel options.

How to use it: Open Unsloth and look in the personalization settings for the Glass styles, greeting, background and code font options. Native glass works in Unsloth Desktop on Mac. The Trick or Treat theme can be used year-round without the Halloween extras. This is release v0.1.914-beta on GitHub.

What this means for you: You can now set a custom background for your chats in Unsloth and move the chat box to the bottom of new chats. A preview of DeepSeek-V4.1-Flash-1017 Dynamic v3.1 GGUFs, a file format for running models locally, is also out.

How to use it: To move the chat box, go to Settings > Chat and turn on Composer at the bottom. Custom backgrounds are in release v0.1.913-beta.

What this means for you: Unsloth is a free app for running and customizing open AI models on your own computer. If you queue several fine-tuning jobs (training a model on your own data), up to 4 compatible ones now train together instead of one after another. The first launch is about 2x faster, and a Halloween theme menu is added. Eleven themes are included, such as Pumpkin, Coven and Ghost Ship.

How to use it: Download Unsloth Desktop for Windows, macOS or Linux. It is free and open source. The Halloween themes are under the Trick or treat row in the sidebar and run until Nov 1. Jobs sent while the GPU is busy queue up in order, and each can be cancelled on its own.

What this means for you: Lets a model running on your own computer operate a web browser by itself, with no one clicking along. In the example, Qwen3.8-27B completed tasks like playing chess and painting on its own.

Why it matters: It comes the same day Unsloth added a Browser use switch to its release, and two days after scheduled tasks arrived on Oct 26, so local models can now both run on a timer and act on websites.

How to use it: Unsloth has a guide at unsloth.ai/docs/basics/browser-use, and the code is on GitHub at github.com/unslothai/unsloth. In the app, the browser's new tab page has a Browser use switch that lets models control the browser.

What this means for you: You can set a prompt to run once, daily, on weekdays, weekly or monthly, and each answer is saved as a new chat. It can use a model on your computer or one from a connected API provider (a service you reach with a key). Tasks run only while Unsloth is open, and without tools or web search.

How to use it: Open Scheduled from the sidebar's More menu. For each task, choose the model and whether to be notified after every run, only on failure, or never. Unsloth Desktop is free and open source, for Windows, macOS and Linux.

What this means for you: Unsloth, an app that runs AI models on your own computer, now has a smaller install that only does chat. You can add training later in Settings. It also runs some very large models on a Mac with less memory than they would normally need, though replies are slower. For example, Qwen3.8-Flash-Next fits on a 32 GB Mac.

How to use it: When you first launch Unsloth Desktop, pick Chat only (GGUF) for the smaller install. Add training later in Settings. New models you can run locally include Flint, MOSS-Image and NVIDIA Nemotron 3.5 Bolt E4B.

What this means for you: Unsloth is a local app for running and training open AI models on your own computer. This release lets it run DeepSeek-V4.1-Flash-1017 and the FLUX.3 Image Klein image model, which makes and edits images with up to 10 reference images and runs on 24 GB of graphics memory. Chat can also make images and videos using an image or video model loaded alongside your chat model.

How to use it: To keep image, video and speech models loaded next to your chat model, go to Settings > Resources > Model memory and pick When they fit or Always. This works on computers with a dedicated GPU. To connect a coding agent, run `unsloth mcp add claude-code` (or `codex`, or `cursor`). DeepSeek-V4.1-Flash-1017 needs about 125 GB of memory at 1-bit.

What this means for you: If you run AI models on your own computer with Unsloth, a PC with NVIDIA RTX Spark can now train diffusion (image), decision and multimodal (text plus images) models about 2.7 times faster. Everything stays on your machine, which has up to 128GB of unified memory, a single shared pool of memory.

Why it matters: It comes a day after Unsloth added NVIDIA Nemotron 3.5 and Meta Muse Spark Open on Oct 15, and builds on the local decision model training it added on Oct 7.

How to use it: Use Unsloth on a PC with NVIDIA RTX Spark. The speedup comes from optimized kernels and binaries for CUDA on ARM, which are built into Unsloth.

What this means for you: Unsloth can now run NVIDIA's Nemotron 3.5 models and Meta's Muse Spark Open on your own computer. Nemotron 3.5 Bolt needs about 24 GB of memory, Super VL about 128 GB at 4-bit, and Muse Spark Open about 256 GB at 2-bit. When Claude Code, Codex or OpenCode runs on a local model, follow-up replies re-read much less of the conversation, and a new Claude Code session's first reply starts up to 4x sooner.

How to use it: This is Unsloth release v0.1.908-beta, with a guide for each model. To see the cursor where a model clicks and types in Browser use, go to Settings > Browser > Browser use > Show cursor.

What this means for you: Unsloth, a free app that runs AI models on your own computer, got faster, especially when you have many chats open (up to 10 times faster) and when models answer (up to 5 times faster). You can drag files into Desktop chat again, including more Office documents and EPUBs. You can also make short videos with the LTX-2.5 and Wan2.2 models, from text or from a starting image.

Why it matters: It comes a day after Unsloth added Browser use on Oct 12, which lets local models open pages, click, type and scroll. This release adds browser scripts that run only with your approval each time, or never if you set Run scripts to Never in Settings > Browser.

How to use it: Update to release v0.1.907-beta from github.com/unslothai/unsloth. For script control, go to Settings > Browser. Video generation is on the Images page in the app. Type /rewind in a chat to restore an earlier point in the conversation.

What this means for you: Browser use lets a model running on your own computer open web pages, read them, click, type and scroll in its own tab beside the chat while you watch. It cannot see your other tabs. Unsloth asks before the model signs in, posts a form or presses a button that buys, sends or deletes, and the model never fills in passwords, card numbers or one-time codes. The release also adds Qwen-Image-2.1-Turbo, which makes images in 8 steps.

How to use it: In chat, open the + menu and turn on Browser use, then ask your model to look something up or use a website. Choose which sites models can open in Settings > Browser. Qwen-Image-2.1-Turbo runs from the Images page with a 4.2 GB Q4 GGUF file. Unsloth Desktop is free and open source for Windows, macOS and Linux.