United StatesMeta’s new open model scores 39, behind 4 Chinese rivals.
Meta calls it America’s strongest open-weight model, yet Chinese models lead on Artificial Analysis’s index.
Meta says the 816-billion-parameter model activates 42 billion per token and runs across two Mac Studios sharing about 1TB of memory.
On the DeepSWE 1.1 coding test it scores 60, below GLM 5.3’s 67 and Kimi K3’s 68.
Independent testers have yet to show whether it finishes complex real tasks reliably.
How each outlet framed it
drawn from 60+ reports worldwide
- Zhidx
- ranks Muse Spark Open's 39-point Artificial Analysis score below MiMo, GLM and Ling-3.1-flash; coding trails Chinese models
- mint
- adds Meta's Helion GPU kernels in SGLang, claimed 1.03x–3.95x speedups on Nvidia and AMD hardware