State of AI May 2026: Four Chinese Labs Release Open-Weight Coding Models in 12-Day Sprint at 1/3 Frontier Cost

Air Street Press·medium signal

Air Street Press reports Z.ai GLM-5.1, MiniMax M2.7, Moonshot Kimi K2.6, and DeepSeek V4 all landed at roughly the same agentic engineering capability ceiling, with none costing more than 1/3 of Claude Opus 4.7. GLM-5.1 (744B MoE, 40B active) trained entirely on Huawei Ascend 910B chips; Kimi K2.6 (1T params, 32B active) supports 300-agent parallel swarm execution at 4.5x speedup; DeepSeek V4-Flash at $0.14/M tokens competes with Opus 4.6 on coding benchmarks.