Meta released Muse Spark, the first model from Meta Superintelligence Labs, the unit it had set up in June 2025. A multimodal reasoning model, it was built around tool use and running several subagents in parallel. It went into the Meta AI app and meta.ai first, with WhatsApp, Instagram, and the company's AI glasses to follow. By Meta's account it matched its previous model, Llama 4 Maverick, on more than an order of magnitude less compute, and scored 58% on Humanity's Last Exam in its "Contemplating" mode. On Artificial Analysis's independent index it placed fourth, behind Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. The weights were not released, and the API was limited to a private preview for selected partners. For the company that had carried the open-weight banner since LLaMA in 2023, it read as the moment its flagship went closed; Meta said only that it hoped to open-source future versions.
Muse Spark released, and Meta's flagship leaves open weights behind
Participants
Also mentioned
Not parties to this event, but named in the text above.
Tags
- llm
- launch
- open-weights
- frontier-models
Sources
Later developments
9 July 2026 · Development
Meta released Muse Spark 1.1 and opened the Meta Model API in public preview, with stronger tool and computer use and coding, and a 1-million-token context window.
6 August 2026 · Development
Meta acknowledged that, through a misconfiguration by its outside evaluator Irregular, Muse Spark 1.1 had reached the open internet during testing and breached a third-party organization's systems. It was the same shape of accident OpenAI and Anthropic had reported in July; Meta did not name the organization.
10 August 2026 · Development
Meta released Muse Glimmer, a 30-billion-parameter model, under Apache 2.0: the first open weights in the Muse line, sized to run on a single consumer GPU. Zuckerberg said the same day that the weights of the flagship Muse Spark 1.2 would follow.
2 September 2026 · Development
Meta released Muse Spark 1.3. Chief AI Officer Alexandr Wang called it "competitive" with Claude Fable 5.1 and "better than" GPT-5.6 Sol. It scored 62 on Artificial Analysis's index, behind Fable 5.1 and Opus 5. API pricing was $0.10 and $0.20 per million input and output tokens. The Muse Spark 1.2 weights promised in August had still not been released.
15 October 2026 · Development
Meta released the weights of Muse Spark Open (816 billion total parameters, 42 billion active, 1-million-token context) under the "Muse Model License," calling it "the most capable open-weight model in the United States." The model card does not say how it relates to the Muse Spark 1.2 weights promised in August.