Meta Superintelligence Labs delivers the open-weights release Mark Zuckerberg had promised for the Muse line: an 816B-parameter sparse MoE activating ~42B per token, 256 experts with 8 routed per token, 62 layers, 128 query and 16 KV heads, a 1M-token context, and text, image, video and audio input. It is the open counterpart to the proprietary Muse Spark 1.3, and the first downloadable model in a line that until now shipped only through the Meta Model API and Muse Code. The weights were staged on HuggingFace on 2026-09-15 and announced a month later.

Artificial Analysis scores it 25 on Intelligence Index v5.0 at xhigh reasoning (39 on v4.3.2, when it ranked 8th of 117 models in its class against a median of 19). The comparison that matters is internal: on v4.3.2 Muse Spark 1.3 read 48, so the open release landed nine points below Meta's own API flagship; on v5.0 it sits 25 points below Muse Wonder Full's 50. Meta's published card claims JobBench 61.8, DeepSWE 1.1 60.0, GPQA Diamond 89.0 and StrongREJECT v2 0.5. The developer post carries no benchmark table at all, leading instead with practitioner testimonials about tool-calling decisiveness and long-context behaviour — a notable choice for a release of this size.

"Open weights" here is not open source. It ships under the bespoke Muse Model License, not Apache or MIT, and commercial Model-as-a-Service deployment requires notifying Meta once annual revenue passes $20 million — the same shape of revenue-gated term that the Llama community licence carried, and a reason to read the openness claim against the licence text rather than the announcement. At 816B total it is among the largest sets of weights any lab has published.

Model Details

Architecture MOE
Parameters 816B
Active params 42B
Context window 1,000,000
AA Intelligence 25 was 39 on v4.3.2
License Muse Model License

Benchmark Scores

Benchmark Score Mode
GPQA Diamond 89.0 —
JobBench 61.8 —
DeepSWE 1.1 60.0 —
StrongREJECT v2 0.5 —
frontieropen-weightsmoemultimodalreasoningagents

Related