Muse Spark

2026-04-08 · Publisher : Meta · Domain : Artificial Intelligence

Meta Superintelligence Labs présente Muse Spark, son premier modèle multimodal de raisonnement, proposé sur meta.ai et l'application Meta AI sans poids ouverts, en rupture avec la ligne Llama.

Meta Superintelligence Labs introduces Muse Spark, its first multimodal reasoning model, offered on meta.ai and the Meta AI app without open weights, a break from the Llama line. ### Context Llama 4 (April 2025) had disappointed and Behemoth was never released. After the July 2025 letter, Meta rebuilt its training stack within Meta Superintelligence Labs. In late 2025, the press mentioned a closed model named Avocado, but no official announcement of that name was found; the commercial name chosen is Muse Spark. ### What the model brings - Native multimodal reasoning, with tool use, visual chain of thought and orchestration of sub-agents in parallel. - “Contemplating” mode, which has several agents reason in parallel; Meta claims 58% on Humanity's Last Exam and 38% on FrontierScience Research in this mode. - According to Meta, the same capabilities as Llama 4 Maverick with more than an order of magnitude less compute for training. - Health expertise developed with more than 1,000 physicians (Meta announcement). - Safety evaluation under the Advanced AI Scaling Framework v2; Meta reports a strong evaluation awareness in the model, to be investigated further. ### Architecture and specifications Parameters, number of experts, context, knowledge cutoff: not disclosed. Modalities: text and image input, text output (Artificial Analysis). Availability: meta.ai and the Meta AI app, with a gradual rollout to WhatsApp, Instagram, Facebook, Messenger and glasses; private API preview for partners. License: proprietary, weights not published; Meta writes that it hopes to open future versions. ### Cost and access | Item | Value | |---|---| | API pricing | Pricing not disclosed at launch (private preview) | | Consumer access | meta.ai, Meta AI app | | Self-hosting | Not possible (closed weights) | ### Benchmarks | Benchmark (Artificial Analysis) | Muse Spark | Gemini 3.1 Pro | Others | |---|---|---|---| | Intelligence Index | 52 | 57 | GPT-5.4 57; Claude Opus 4.6 53 | | Humanity's Last Exam | 39.9% | 44.7% | not disclosed | | MMMU-Pro | 80.5% | 82.4% | not disclosed | | GDPval-AA | 1,427 | not disclosed | GPT-5.4 1,676; Claude Sonnet 4.6 1,648 | Source: Artificial Analysis (third party), April 2026. Output tokens consumed for the index: 58M versus 157M for Claude Opus 4.6 and 120M for GPT-5.4. ### Versus closed models Muse Spark is itself a closed model, so the comparison is with its proprietary peers. According to Artificial Analysis, it ranks fourth on the index with 52 points, behind Gemini 3.1 Pro (57), GPT-5.4 (57) and Claude Opus 4.6 (53). On agentic tasks (GDPval-AA), it remains clearly behind GPT-5.4 and Claude Sonnet 4.6. Artificial Analysis presents it as the first Meta model not released with open weights and notes a jump since Llama 4 Maverick (18 on the current version of the index). For an open-versus-closed choice, there is no self-hosting and no license: access depends on a Meta API that is still in preview. ### Key takeaways Muse Spark confirms that Meta's frontier now runs through closed weights. The catch-up with the leaders is real according to a third party, but cost and context data were missing at launch.

Tags : LLM, Meta, Muse, Poids fermés, Multimodal, Raisonnement, Meta Superintelligence Labs