Muse Wonder raises Meta's ceiling, llm-meta-ai 0.2 brings it to the terminal

Simon Willison released llm-meta-ai 0.2, a plugin for using the Meta AI API through his LLM tool. The update supports the newly released Muse Wonder and Muse Wonder Full models and gives developers a short route to testing them.

The plugin exposes six reasoning levels

Willison showed Muse Wonder outputs for minimal, low, medium, high, xhigh and max reasoning. He used his recurring image prompt of a pelican riding a bicycle and published the highest setting alongside earlier results. Meta describes the two models as its most capable models.

The comparison also includes Muse 1.2 from August 5 and Muse 1.3 from September 2. Three versions in roughly two months show a fast release cadence. The images are a useful visual check, not a standardized benchmark.

Terminal access shortens the path from launch to a private eval

The plugin matters most to people who do not want to judge a model from a vendor demo. A team can run the same prompt across reasoning levels and keep the results beside responses from other models in one tool.

That changes the first hours after a launch. Instead of waiting for a large platform to add support, developers can run a small eval on their own tasks and see whether more reasoning improves the answer or merely extends the run.

A pelican shows progress but cannot measure production work

A repeated image prompt can reveal problems with anatomy, composition and instruction following. It says little about code accuracy, long context work or API reliability. Artificial Analysis scoring Muse Wonder Max around GPT-6.1 Sol is another useful signal, but it cannot replace a test on the workflow that will actually use the model.

The six settings may also trade quality against latency and price. Without measuring all three, max is the strongest demonstration rather than an automatic production default.

Real tasks will decide whether the monthly cadence matters

The next useful evidence will be repeatable evaluations of Muse Wonder against Muse 1.3 with identical prompts, settings and budgets. Users need the distribution of results and response times, not only the best sample.

llm-meta-ai 0.2 removes one obstacle to that work. If the improvement survives tests on code, data extraction and tool use, Meta's release cadence will mean more than a regular change of model names.

Lilith's verdict

A pelican on a bicycle is a useful smoke test, but a production model earns its place when a team hands it broken code from its own stack. llm-meta-ai 0.2 gets people into that testing room without a ribbon ceremony.