Muse Spark 1.1 and Meta Model API

2026-07-09 · Publisher : Meta · Domain : Artificial Intelligence

Meta launches Muse Spark 1.1 in its Meta Model API as a public preview: a multimodal reasoning model with a one-million-token context, oriented toward agents, tools and code.

What changes

  • Gains announced by Meta in tool and computer use, code and multimodal understanding.
  • 1-million-token window with active context management (Meta).
  • Access through the Meta Model API in public preview, with an OpenAI-compatible package cited by a partner.
  • "Thinking" mode in the Meta AI app and on meta.ai.

### Context Muse Spark (April 2026) was only available in an app and as a private API preview. Meta wanted to give developers direct access to compete with Claude, GPT and Gemini on agentic use cases. ### What the model brings - Gains announced by Meta in tool and computer use, code and multimodal understanding. - 1-million-token window with active context management (Meta). - Access through the Meta Model API in public preview, with an OpenAI-compatible package cited by a partner. - "Thinking" mode in the Meta AI app and on meta.ai. ### Architecture and specifications Size, architecture, data and knowledge cutoff: not disclosed. Context: 1 M tokens. Modalities: text and image input (to be confirmed on the API side). License: proprietary, weights not published. API identifier: not read on the page. ### Cost and access | Item | Value | |---|---| | Pricing | Pricing not disclosed in the post read (tiers appear with version 1.2) | | Access | Meta Model API (public preview), Meta AI app | ### Benchmarks The post refers to evaluation charts whose values could not be extracted. Third-party reference: according to Artificial Analysis, Muse Spark 1.1 scores 53 points on its Intelligence Index, versus 57 for Muse Spark 1.2 and 61 for Muse Spark 1.3 (xhigh) in the same comparison; other sources give different values for 1.2, because of index versions. ### Versus closed models In the text read, Meta does not publish a numerical table against closed models. The third party Artificial Analysis places version 1.1 at 53 points, at the level of a leading model a quarter earlier (Claude Opus 4.6 at 53 on the April index, index versions not necessarily identical). No price is cited, so the cost gap cannot be established for this version. The only firm element is public API access, which makes the model testable in the enterprise without a private contract. It remains without open weights or self-hosting. ### Key takeaways Muse Spark 1.1 turns the Muse family into a developer product, with a context of one million tokens. Detailed official pricing and scores await version 1.2.

Links

Tags : LLM, Meta, Muse, Poids fermés, API, Agents, Multimodal