Vibeleaderboard
← All Intel
Intel / repo

unsloth v0.1.919-beta — Computer Use + Translate + MiniMax-H3-Image + New Model Picker

Source
github.com
Author
danielhanchen
Date

Context

Unsloth, whose open-source software trains and runs AI models on a person's own hardware, now lets models operate the computer in its app. In this computer use beta, a model looks at screenshots of the screen and controls the mouse and keyboard on macOS, Windows and Linux. It works with vision models running locally in GGUF, safetensors or MLX format, which are file formats for running models on your own machine, and with connected cloud providers. The feature is off by default. According to Unsloth's computer use guide, the app has to be installed directly on the computer it controls, since Docker and WSL cannot reach the desktop, and the model must handle both images and tool calls. On Linux it needs an X11 session.

The safeguards are layered. The first action in each chat needs approval unless the user has chosen Full access mode, and the first click or keystroke in each new app asks for permission again. Steps the guide treats as risky, such as opening a terminal or password manager, using system tools, or shortcuts that close or delete, always ask. Apps on the blocked list are never touched in any mode, and the model can never type into operating-system security prompts or press keys that lock the screen or end the session. The user can stop it with the Stop button or by moving the pointer into any corner of the screen.

Unsloth's own tests show how early this is. The best performer in the guide was Qwen3.8-27B, about 19 GB including its vision file, while small models such as gemma-4-E4B could not point at buttons at all. Unsloth ran a sample task, reading six receipt photos into a spreadsheet with a total and a chart and turning a company template into a three-slide summary, six times with Qwen3.8-27B on one A100 GPU. Two runs got everything right, taking 22 and 30 minutes. Two misread an amount, one stopped before making the deck, and one charted the wrong column, a failure the guide traces to an earlier version of its prompt. The guide tells users to check results as they would a colleague's and to stop the model if it turns to a terminal or developer tools. Running a screen-controlling entirely on local hardware is now possible in Unsloth's app, but its own results describe a tool to supervise, not one to leave alone with a task.

Key takeaways · AI-distilled
  • Unsloth's computer use is off by default and asks for approval before the first action, before switching apps and before potentially risky actions; blocked apps stay inaccessible and the user can take control at any time.
  • MiniMax-H3-Image GGUF builds run through stable-diffusion.cpp on CPUs, Macs, smaller GPUs and AMD, and automatic precision picks INT8, FP8 or BF16 by hardware; Unsloth reports a 2048px decode on an A100 80GB dropping from 2.82s to 0.28s.
  • Qwen3.8-Flash-Next and DeepSeek-V4.1 GGUFs now store their Engram table in a separate file, so any quant can be paired with a smaller or lossless table, chosen in advanced settings or with an --engram flag on the command line.
  • Long chats on connected cloud providers are now summarized automatically rather than dropping older messages, with limits configurable per connection and /compact working with cloud models.
  • On the training side, Unsloth says new kernels make Gemma 4 12B training 1.4 to 2.6 times faster, and SFT now packs text datasets by default on epoch-based runs.
Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • 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…