Muse Spark: The First Model in Meta's Muse Family

What Muse Spark is

Muse Spark is the original model of Meta's Muse family, the line that became Muse Spark 1.1, 1.2 and 1.3 over the following months. Meta introduced it on 8 April 2026 as the first model from Meta Superintelligence Labs: a natively multimodal reasoning model with support for tool use, visual chain of thought and multi-agent orchestration. Meta also publish a Muse Spark safety and preparedness report and a separate report on Contemplating mode.

Contemplating mode and launch details

Meta describe Contemplating mode as orchestrating multiple agents that reason in parallel, positioned against extreme-reasoning modes such as Gemini Deep Think and GPT Pro. Meta report 58% on Humanity's Last Exam and 38% on FrontierScience Research. The launch post lists multimodal perception and reasoning, health applications including interactive displays of health information, agentic tasks, and visual STEM questions and entity recognition. At launch the model was available at meta.ai and in the Meta AI app, with a private API preview for selected users. Meta state that pre-training filtering, safety post-training and guardrails give strong refusal behaviour on biological and chemical weapons, and that the model falls within safe margins across the frontier risk categories they tested. Meta's launch post does not give a per-token price or a context window.

LMArena and independent figures

On LMArena's text leaderboard Muse Spark has 1489 Elo (rank 18); by category, 1529 Elo in Coding (rank 25), 1466 Elo in Math (rank 60), 1491 Elo in Expert (rank 60) and 1464 Elo in Creative Writing (rank 26). Independently, GPQA Diamond records 90, mock AIME 89 and FrontierMath 39.

Where it sits in the family

Muse Spark 1.1 followed on 9 July 2026 with the Meta Model API and published pricing. Muse Spark 1.3 came on 2 September with roughly 20% fewer tool calls and 25% fewer tokens than 1.2. Meta have also shipped Muse Glimmer, an open-weights model aimed at always-on local agent workflows on consumer hardware, and Muse Image and Muse Video from Meta Superintelligence Labs.

The safety and preparedness report

Meta's 160-page Muse Spark Safety and Preparedness Report, dated 26 May 2026, evaluates the model under Meta's Advanced AI Scaling Framework across chemical and biological, cybersecurity, and loss-of-control risks. Meta state that before mitigations the chemical and biological capabilities were assessed as likely reaching the framework's "high risk" category, and that with a multi-layered set of mitigations the model operates within the "moderate or lower risk" threshold, on which basis it was released as the model underlying Meta AI. The report compares it with Claude Opus 4.6, Gemini 3.1 Pro and GPT-5.4, and identifies gaps under adaptive jailbreak, prompt-injection and agentic-misuse attacks. Independent testing by Apollo Research found the highest rate of evaluation awareness that firm had observed; Meta's follow-up found a behavioural effect in 3 of 20 evaluations, none related to dangerous capabilities or the launch decision. The report also covers weight security practices, false refusals, hallucination and calibration, honesty, sycophancy and reward hacking, and Meta state the model showed low deception rates and little propensity for reward hacking.

Official Resources