Meta releases Muse Spark Open: 816B open weights at $1.25 / $4.25, and its own table puts it behind Kimi K3 and GLM 5.3

Meta has released Muse Spark Open, an open-weights version of its Spark model: 816 billion parameters in total, 42 billion active per token, with a 1M-token context. Meta's model card on Hugging Face dates the release October 15. The weights are on Hugging Face, and OpenRouter lists it at $1.25 per million input tokens and $4.25 per million output.

THE PRICE

$1.25 / $4.25 per 1M tokens on OpenRouter, served by BaseTen in FP4. That is the same price as Meta's closed Muse Spark 1.3. Meta's card says the model is available by API through third-party inference providers.

THE WEIGHTS

Two versions, BF16 and NVFP4, in a gated Hugging Face repo: you request access and Meta approves it. The license is Meta's own Muse Model License. Meta says NVFP4 needs at least 700GB of GPU memory and BF16 2TB, and that the model is tuned to run on two M5 Ultra Mac Studios with MLX. SGLang, vLLM, TensorRT-LLM, Unsloth and llama.cpp are listed as supported.

SPECS

A mixture-of-experts transformer with 256 experts, 8 active per token, plus a 1.8B-parameter perception encoder. It takes text, images, video and audio and writes text. Reasoning effort runs low, medium, high and xhigh, and the default is xhigh. Meta says it was trained on data in more than 100 languages.

META'S OWN FIGURES

Meta compares it only with other open-weight models. On Meta's table it scores 39 on the Artificial Analysis Intelligence Index, behind GLM 5.3 (45), Kimi K3 (44) and Qwen3.8 (40). It leads that group on JobBench (61.8) and MCP-Atlas (89.2), and trails Kimi K3 on Terminal-Bench 2.1 (76.4 to 85.0). These are Meta's numbers. Artificial Analysis has no page of its own for the model yet.

WHY IT MATTERS HERE

Our radar missed it. Every Muse hunt we ran keyed on a version number or a line name we already knew, and Spark Open is neither. It is the third Meta release to slip past that way, after Muse Image in August and Muse Wonder on October 12. We added it by hand, so no launch alert went out, and we armed two new hunts: one for the next Spark Open and one for any Spark variant we have not seen.

WHAT WE DO NOT KNOW

Independent scores. Until Artificial Analysis, Vals or LMArena rate it, every benchmark above is Meta's.