Vibeleaderboard
← All Intel
Intel / repo

unsloth v0.1.918-beta — Computer Use, MiniMax-H3-Image + Perf fixes

Source
github.com
Author
danielhanchen
Date

Context

The newest release of Unsloth's app, Unsloth Studio, adds a beta feature called computer use: an AI model can look at your screen and operate the mouse and keyboard on macOS, Windows and Linux, the way a person sitting at the computer would. According to the v0.1.918-beta release notes, it works with vision models, meaning models that can read images as well as text, run on your own machine in the GGUF, safetensors or MLX formats, and with models from cloud providers connected to the app. A model that only writes text needs each program to offer it a dedicated connection; a model that can read the screen and click can work through a program's ordinary interface instead.

Handing a model your mouse and keyboard is a bigger grant of trust than letting it answer questions, and the release notes describe several limits on it. The feature is off by default. It asks for your approval before its first action, before it switches from one app to another, and before potentially risky actions, though the notes do not say which actions count as risky. Apps you block stay inaccessible to it, and you can take back control at any time.

The same release widens a second, narrower route for models to act on other software. Unity, Unreal Engine, Godot, Figma and Roblox now appear in Studio's menu of connections alongside Blender. MCP, the Model Context Protocol, is a standard way for an application to offer a model a defined set of actions, so these connections let a model build scenes and work on game and design projects through each program's own interface rather than through the screen. Figma's connection is read-only. Earlier this month, Unsloth's v0.1.915-beta release let models interact with files in Studio's built-in browser; computer use extends that kind of control from inside the app to the rest of the desktop.

What the release establishes is the capability and its permission controls, not how well it works. The notes report no success rates, test results or examples of tasks completed through computer use, and they do not compare how different models perform with it. Until Unsloth or users publish results, it is best read as a beta to experiment with, with the approval prompts and blocked-app list as its main protection.

Key takeaways · AI-distilled
  • Unsloth's computer use beta lets local GGUF, safetensors and MLX vision models, or connected providers, see the screen and drive mouse and keyboard on macOS, Windows and Linux. It is off by default and asks for approval before the first action, app switches and risky actions.
  • MiniMax-H3-Image now runs as GGUF through stable-diffusion.cpp on CPUs, Macs and AMD, with FP8 and INT8 weights for NVIDIA. Unsloth reports a 2048px image decoding in 0.28s instead of 2.82s on an A100 80GB.
  • Qwen3.8-Flash-Next and DeepSeek-V4.1 GGUFs now keep their Engram table in a separate file, so users can pair any quant with a smaller or lossless table, chosen in the model picker or with the --engram flag.
  • Training changes include 1.4 to 2.6x faster Gemma 4 12B training from new kernels, native , QLoRA and distributed training for DeepSeek-V4.1, and sequence packing on by default for epoch-based SFT runs.
  • Long conversations on connected providers are now automatically summarized instead of losing older messages, with limits set per connection, and /compact now works with cloud models.
Terms in this piece · Glossary
  • MCP — The Model Context Protocol — an open standard that lets any AI assistant plug into any tool or data source without custom integration code.
  • LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.
  • context compaction — Summarizing an agent's earlier conversation to free room in the context window so a long session can keep going.
  • attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
More from danielhanchen
Recommended reads
Comments

Checking sign-in…

Loading comments…