Screen access comes with stated approval boundaries

According to the release, computer use supports macOS, Windows and Linux, local vision models in GGUF, safetensors and MLX formats, and connected providers. It is disabled by default. The default workflow asks before the first action, when switching applications and for potentially risky actions. The documentation qualifies this: Full access exempts the initial approval, and individual applications can be preapproved. Blocked applications remain inaccessible, and the user can take control. These are the team’s described safeguards, not findings from an independent security audit. A compatible vision model is still necessary: supporting a weight format does not by itself give a text-only model the ability to understand screenshots.

Application tools and trained adapters solve different problems

The MCP menu gains Unity, Unreal Engine, Godot, Roblox and Figma alongside Blender. Figma access is described as read-only. An integration exposes tools for working with a project or scene; it is a different form of specialization from fine-tuning. MCP connects application capabilities, while LoRA changes model behavior through trained adapter weights. Whether a setup can usefully build a game or edit a scene depends on the tools exposed, the permissions granted and the underlying model. The notes do not establish a general success rate across these creative workflows.

Image formats expand the hardware choices

MiniMax-H3-Image gains GGUF support through stable-diffusion.cpp for CPUs, Macs and AMD hardware, plus FP8 and INT8 weights for NVIDIA GPUs. Training and loading custom LoRAs are also listed. Format and precision choices help match deployment to hardware, but do not replace checking the relevant model card, license and memory requirements. The team reports that decoding a 2048-pixel image takes 0.28 seconds instead of 2.82 on an A100 80GB. That is a measurement of one stage on specified hardware, rather than an end-to-end image-generation guarantee for every device.

Connections and conversation compaction change the workflow

The revised model picker brings local and connected models together. Connections displays live status and errors such as exhausted credit or rate limits. Long conversations with connected providers can now be summarized automatically, with configurable thresholds. A summary is not an exact preservation of the full conversation, so important task requirements may need checking after compaction. For task-specific deployments, this release offers several separate controls: application tools, model choice, numerical precision and adapters. Evaluating those controls independently is more informative than treating computer use as a single capability score, especially while the feature remains in beta.