Summary
Meta launched Muse Spark 1.3, a multimodal frontier model tuned for longer-horizon agentic and coding work that Meta says finishes coding tasks with roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2, with a 1M-token context.
What changed
On September 2, 2026 Meta released Muse Spark 1.3 in Muse Code and the Meta Model API. It reads text, images, and video, holds a million tokens of context, and is trained to sustain multi-workflow threads, ask clarifying questions, call for help when stuck, and confirm before irreversible actions. Meta reports ~20% fewer tool calls and ~25% fewer tokens than 1.2 on coding work; a max-reasoning mode follows after further safety testing.
Why it matters
Efficiency, not just accuracy, is becoming the competitive axis: fewer tool calls and tokens directly cut the cost and latency of agentic coding at scale. It keeps Meta credibly in the frontier coding race against OpenAI, Anthropic, and Google, and its availability via the Meta Model API broadens who can build on it.
Evidence excerpt
Meta AI Research: 'Introducing Muse Spark 1.3' (September 2, 2026). Meta's engineers measured it finishing coding work with roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2; 1M-token context; reads text, images, and video; max-reasoning mode to follow after safety testing.