US open models are back from irrelevance, Lambert argues, but barely

Nathan Lambert argues that two October releases, Meta's Muse Spark Open and Reflection's Beam, have pulled American open-weight models back from the edge of irrelevance. His framing is blunt: without those two models, more than 20 Chinese open-weight models would beat the best Western ones. With them, the gap between American and Chinese open models has roughly halved, moving US open weights from "largely irrelevant" to "behind but relevant."

He ranks the labs in order: Anthropic, then OpenAI, then Meta and Google roughly level, then xAI alongside the Chinese open-weight labs, with US open weights last. Meta's position is unusual because it is now the only American company shipping both a frontier open model and frontier closed ones. Its closed Muse Wonder models, Lambert writes, trail only Anthropic's Opus 5.5 and sit close to OpenAI's GPT-6-Astra and Google's Gemini-4-Argon.

His main worry is the license, not the weights. Muse Spark Open carries non-standard terms: products must display "Built with Muse," companies earning over $20 million a month need a separate Meta license, and the restrictions reach into outputs and derivatives, which muddies whether you can train on synthetic data from it. Lambert calls complex, custom licenses a self-inflicted strategic risk for a lab trying to win back adoption, since Western open models tend to get picked up more than Chinese ones at similar quality. He closes on a market note: Meta is valued under $2 trillion while Anthropic is reportedly nearing an IPO at a similar figure, and he thinks one of those numbers has to move.

Why it matters

If you build on open-weight models, there is again a capable American option on the table, but on licensing Beam's Apache 2.0 is the cleaner bet. Read Muse Spark Open's terms carefully before you commit a product to it.