[KongchangAI]
ModelMuseOpen

MuseSpark Open

An open-weight large model released by Meta, aimed at coding and long-horizon agentic tasks, with emphasis on the competitiveness of open models from the United States.

Timeline (last 90 days)

Oct 17

Meta released MuseSpark Open, calling it the strongest open-weight model from the US to date, with 816 billion total parameters, 42 billion active parameters, a 1-million-token context window, and aimed at coding and long-horizon agentic tasks.

Unverified50%
Oct 17

Artificial Analysis gave MuseSpark Open a score of 39, tied for 17th on the overall leaderboard, 7th among open-source models, about 7 points behind the open-source SOTA, and about 13 points ahead of Nemotron and Inkling.

Unverified50%
Oct 17

Meta claims MuseSpark Open achieves open-source SOTA on agentic benchmarks such as JobBench and MCPAtlas.

Unverified50%
Oct 17

Open-source model researcher Nathan Lambert said of MuseSpark Open that before this, US models still lagged behind Chinese models, and MuseSpark Open has at least halved the gap.

Unverified50%
Oct 17

Apple updated its MLX toolkit to ensure MuseSpark Open runs smoothly on its own hardware, and Zuckerberg posted a video of Muse running on a Mac Studio and thanked Apple's John Ternus.

Unverified50%
Oct 17

The full version of the MuseSpark Open model requires 1TB of memory, roughly equivalent to two fully specced M5 Ultra Mac Studios, priced at about $40,000.

Unverified50%

All Facts (6)

Source Articles