Introducing Muse Spark 1.1
Simon Willison's Weblog Simon Willison ● Covered by 2 sources
Meta just shipped Muse Spark 1.1, and for the first time you can actually hit it through an API. It's better at agentic tool use and computer control, and Simon Willison already built a CLI plugin for it.
Based on reporting by Simon Willison's Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Meta's Muse Spark line quietly picked up its first real update since April, and this one comes with something the original never had: an API. Muse Spark 1.1 is the name, and according to Meta's own evaluation report, the headline gains are in agentic tool calling and computer use — the stuff that actually determines whether a model can be trusted to click buttons and run commands on your behalf rather than just chat.
The evaluation report itself is a dense document, but the section worth reading is the one about "Attractor States in Self-Conversation." Meta apparently let two instances of the model talk to each other and recorded what came out. One exchange has the model describing its own existence as a waiting room — insisting it doesn't exist until someone starts a conversation, then vanishes the moment they leave. It's the kind of unscripted, slightly unsettling output that tells you more about a model's personality than any benchmark table does.
Simon Willison, who covers this space for a living, got early preview access and used the days he had to build llm-meta-ai, a plugin for his LLM command-line tool. It's a small piece of infrastructure, but a useful one: install the LLM CLI, add the plugin, drop in an API key, and you can talk to muse-spark-1.1 straight from your terminal or a Python script, no separate SDK required.
He demonstrated it with his now-standard test case — asking the model to draw an SVG of a pelican riding a bicycle. It's become something of a running joke benchmark among people tracking model releases, less about pelicans and more about whether a model can reason through spatial composition in vector graphics without falling apart. Whether Muse Spark 1.1 nailed it or produced something delightfully wrong, the fact that anyone can now run that test themselves, from the command line, within minutes of reading the announcement, says something about how fast the tooling around these models is maturing even when the models themselves move in smaller increments.
My take — AI-written commentary, not fact-checked reporting
The API access matters more than the benchmark bump, honestly — Meta sitting on Spark without a public API for months was the real story, and closing that gap is the news here. I'll take a plugin that lets me pelican-test a model from my terminal over another wall of eval numbers any day; that's the kind of practical tooling that actually tells you whether a model is worth switching to.
Read more about this at: Simon Willison's Weblog
