US open models closed the gap with China, then added a license problem

For most of this year the best open-weight models have come from Chinese labs. Nathan Lambert argues that changed in the past two weeks. With Meta's Muse Spark Open and Reflection's Beam, he says American open models have roughly halved the gap with the Chinese frontier. His own count is blunt: strip out those two releases and more than twenty Chinese open models would still beat the best Western one.

Muse Spark Open is the centerpiece. At 816 billion total parameters with 42 billion active and a one million token context, Lambert places it above GLM 5.2 and close to DeepSeek's V4.1 Flash, though still behind GLM 5.3 and Kimi K3. Meta is now the only American company shipping both a frontier open model and a frontier closed one, with Muse Wonder trailing only Claude Opus 5.5 on standard benchmarks.

The catch is the license. Meta's terms require a "Built with Muse" credit, a separate license for any company above $20 million in monthly revenue, and they reach into model outputs and derivatives, which muddies whether you can train on synthetic data from the model. Lambert calls this an unnecessary strategic risk: open models from the West already get more real adoption than Chinese ones at similar quality, so adding legal friction works against the one advantage these labs have. He expects most users to ignore the terms anyway.

Why it matters

If you were reaching for a Chinese model because nothing Western competed on open weights, that calculus just shifted, but read Meta's license before you build on Muse. The revenue threshold and the clause covering outputs can decide whether it is usable for your company at all.