Muse Spark 1.1

Muse Spark 1.1 is Meta's multimodal reasoning model aimed squarely at agentic workloads — writing and debugging code, chaining tool calls, and driving a computer through multi-step tasks. Announced on July 9, 2026 alongside a public preview of the Meta Model API, it ships with a 1M-token context window, which is the headline spec for developers who want to feed whole repositories, long logs, or multi-document task histories into a single request. Unusually for a model release, Mark Zuckerberg amplified the launch personally on X, framing coding and agentic performance as the differentiators.

The model sits in Meta's lineup above the general-purpose Llama family, positioned as the reasoning-and-agents tier rather than an open-weights checkpoint you self-host. Everyday users get it free inside the Meta AI app, while developers reach it through the Meta Model API — a notable shift for Meta, whose reputation has been built on open-weight releases rather than a hosted, metered API. That makes Muse Spark 1.1 as much a strategic statement as a technical one: Meta competing directly on the same agentic-coding turf as Anthropic's Claude, OpenAI's GPT-5.x, and Google's Gemini, through a paid endpoint.

What's still unknown matters here. Meta has not published API pricing, and independent benchmarks for the coding and computer-use claims weren't available at launch — the performance narrative so far is Meta's own. For teams evaluating it against established agentic coders, that means treating the launch numbers as vendor-reported until third parties weigh in.