Meta's Muse Spark 1.1 targets agentic coding and computer use with 1M-token context

Meta Superintelligence Labs has released Muse Spark 1.1, a multimodal reasoning model aimed squarely at agentic work — planning, tool use, and computer control rather than plain chat. The headline capabilities are orchestration and coding: the model can act as a lead agent that gathers context, plans, and delegates to parallel subagents, or serve as a subagent that sticks to its assigned task and escalates when needed. It ships with a 1-million-token context window it manages actively, retrieving earlier work and compacting while preserving steps that matter later. On computer-use tasks, Meta says it decides when to script an action versus click through an interface directly, and adapts mid-task as conditions change.

Coding is the clear commercial focus. Meta claims substantial gains on real-world work over large codebases — bug diagnosis, feature implementation, and code migrations — and says the model plugs into popular agentic harnesses with support for planning mode, subagent delegation, and context compaction. Demos highlight self-directed debugging (building an app, screenshotting failures, tracing them to code) and even having the model evaluate itself on a subset of SWE tasks. Launch partners including Replit, Cline, and Box supplied endorsements emphasizing long context, multimodal input, and tool use at a price point viable for large-scale coding workloads.

Alongside the model, Meta is opening a public preview of a new OpenAI-compatible Meta Model API, and Muse Spark 1.1 is live in “Thinking” mode in the Meta AI app and on meta.ai. Meta says it ran safety evaluations under its Advanced AI Scaling Framework across chemical/biological, cybersecurity, and loss-of-control risk categories, reporting the model stays within safe margins and shows improved resistance to jailbreaks, prompt injection, and developer-prompt attacks, plus lower hallucination and sycophancy. As always with vendor-published benchmarks and safety claims, independent verification is still pending.