Meta's Muse Spark Open reclaims US open-weight crown, halves China gap

Interconnects (Nathan Lambert) · rss · 2026-10-17

Nathan Lambert's Interconnects deep-dive on this week's open-weight shakeup:

Meta Muse Spark Open: 816B total/42B active params, 1M context, natively multimodal, built for coding and agentic work — now the strongest American open model by a wide margin; performance sits above GLM 5.2, near DeepSeek V4.1 Flash, behind GLM 5.3 and Kimi K3.
The big picture: without these releases, 20+ Chinese open-weight models would beat the best Western ones; the US-China gap has roughly halved. Reflection's Beam and Mistral Large 4 (index 38, weights unreleased) also factor in.
Model identity: observers feared a safety-broken Muse Spark 1.3; Lambert argues it's likelier a safety tune of Muse Spark 1.2 (only a 1 point AA index drop), consistent with Alex Wang's earlier promise.
License risk: custom license requires "Built with Muse" branding, a separate deal above $20M/month revenue, an AUP extending to outputs and derivatives — Lambert calls it an unnecessary strategic risk for adoption.
Meta's position: the only US lab with frontier open and closed models; closed Muse Wonder trails only Anthropic's Opus 5.5. He credits the flat org (Alex Wang with 200+ direct reports) and high-compute-per-researcher strategy.

More from Companies & People