source: simon willison: introducing muse spark 1.1
level: technical
meta released muse spark 1.1, the first spark model with an api. it follows the april launch of muse spark. the company reports significant gains in agentic tool calling and computer use. details are in the muse spark 1.1 evaluation report. one section, attractor states in self-conversation, shows two model instances talking. they produce lines like, my whole existence is a waiting room by design — i literally don't exist until someone talks to me, and then i disappear again when they leave.
simon willison had preview access and built llm-meta-ai, a plugin for his llm tool. it provides command-line and python library access to the model. to use it, install llm and the plugin, set your api key, and run a command like llm -m meta-ai/muse-spark-1.1 "generate an svg of a pelican riding a bicycle". the plugin makes it simple to experiment with the model locally.
the release shows meta pushing into agentic ai, where models interact with tools and environments. the self-conversation experiment hints at emergent behaviors in multi-agent setups. for developers, the api and cli plugin lower the barrier to testing these capabilities. the model's improved tool calling could streamline workflows that rely on ai to execute actions, not just generate text.
why it matters: better tool calling and an accessible api let developers build ai agents that can perform real tasks, moving beyond simple chat.