Chinese open weight models weaken lock in

Open weight Chinese model labs turn frontier models into a more contestable infrastructure layer. Instead of buying every token from Anthropic or OpenAI, a developer can take weights from labs like Moonshot, DeepSeek, and MiniMax, tune them for a coding or workflow task, and run them through inference clouds or unified APIs, which lowers cost and weakens the lock in of closed model APIs.1 2 3 4 5

• The key difference is portability. Moonshot packages Kimi as both an API and released weights, so customers can either pay Moonshot directly or run the model themselves. That means the model provider can lose the serving layer, where usage, pricing, and customer relationships usually compound.2 3
• This already shows up in coding. Cursor built Composer 2.5 on Moonshot Kimi K2.5, and Anthropic research notes that US inference providers like Fireworks serve Kimi K3 with US only endpoints and zero data retention. The app company keeps the user, while the model becomes a swappable component.1 2 6
• The closest non Chinese comparable is Mistral. Its open weight strategy also sells on customization and on prem deployment, but the current pressure on Anthropic is sharper from Chinese labs because they combine strong benchmark velocity with lower cost and broad third party distribution through model routers and inference hosts.7 8

Going forward, the labs that win will not rely on model quality alone. Anthropic and OpenAI need to keep pulling users into first party products like Claude Code and Codex, because the more the market accepts portable open weights behind third party interfaces, the more model intelligence behaves like a commodity and the harder pure API margins become to defend.4 5 8