Four Chinese Labs Just Hit Frontier-Class Coding in 12 Days

DeepSeek V4, Z.ai's GLM-5.1, MiniMax M2.7, and Moonshot's Kimi K2.6 all shipped open-weights coding models within a 12-day window in late April and early May, per llm-stats.com. All four hit roughly the same capability ceiling on agentic engineering benchmarks. All four ship at meaningfully lower inference cost than the Western frontier. DeepSeek-V4-Flash activates only 13 billion parameters per token via a 284-billion-parameter Mixture-of-Experts architecture.

Cost per token is the change, since Western labs hit similar capability last year. The MoE architecture means inference compute scales with active parameters, and the total parameter count drops out of the bill. At 13B active per token against frontier-class output, DeepSeek-V4 is delivering roughly 5–10× lower inference cost than dense models in the same capability tier. The other three Chinese releases use similar tricks. For self-hosting builders, open weights are now economically competitive with API access to Claude or GPT-5.5 on coding workloads, assuming you have the GPUs to run them. The Western labs already match the capability. Their open question is whether to match the cost structure.