Muse Spark 1.1: Meta's Agentic Model and the Meta Model API

What Muse Spark 1.1 is

Muse Spark 1.1 is Meta's agentic language model, introduced on 9 July 2026 and the release that came with an API rather than just an app. It is natively multimodal — images, video and PDFs alongside text — with a 1 million token context window, built-in search with citations, structured output and parallel tool calling. Meta position it as part of what they call personal superintelligence. It runs in Thinking mode in the Meta AI app and on meta.ai, and through the Meta Model API, which launched in public preview alongside it.

The Meta Model API

Until this release Meta's frontier models were something you used in Meta's own products. The Meta Model API is OpenAI-compatible — existing client code and tooling point at it with a base-URL change rather than a rewrite. Pricing is $1.25 per million input tokens and $4.25 per million output. At launch the preview was limited to developers in the United States. One of Meta's quoted assessments lists a million-token context, full multimodal support, search with citations, strong reasoning, coding strength particularly in frontend and design, structured output and parallel tool calling, in a clean OpenAI-compatible package.

Safety and robustness, in Meta's own words

Meta published more on adversarial robustness than most vendors do at launch. Across the frontier risk categories they evaluate — chemical and biological, cybersecurity, and loss of control — they report the model operating within safe margins. They specifically claim strong resistance to direct jailbreaks and to indirect attacks arriving through untrusted data, prompt injection and developer-prompt attacks, and report lower hallucination rates and reduced sycophancy as a consequence. These are Meta's own evaluations on Meta's own harness; no independent replication is published alongside them. Box, testing against their enterprise work evaluation set, described the model as competitive with leading frontier models and singled out structured, procedural workflows in professional services, public sector and industrial operations.

LMArena

On LMArena's text leaderboard Muse Spark 1.1 has 1490 Elo (rank 16); by category, 1533 Elo in Coding (rank 19), 1494 Elo in Math (rank 25), 1503 Elo in Expert (rank 39) and 1444 Elo in Creative Writing (rank 52). SimpleQA Verified records 58.

Licence

The model is proprietary and not self-hostable. Muse Glimmer, released the following month, is the open-weight model in this family.

Agentic and coding features

Meta describe Muse Spark 1.1 as a multimodal reasoning model from Meta Superintelligence Labs, released in the same week as Muse Image. As a main agent it gathers context, makes a plan and delegates to parallel subagents; as a subagent it keeps to its job and escalates back when needed. Meta state it zero-shot generalises to new native tools, MCP servers and custom skills, actively manages its 1 million token context by compacting while keeping critical steps, and in computer use chooses between writing scripts, clicking directly and issuing batches of actions. It supports agentic coding setups with planning mode, goal conditioning, subagent delegation and context compaction. Safety evaluations followed Meta's Advanced AI Scaling Framework and are documented in the Muse Spark 1.1 Evaluation Report.

Official Resources