Four Chinese Open-Weight Coding Models in 12 Days: DeepSeek V4, Kimi K2.6, and the New AI Cost Floor

Between April 7 and April 24, four Chinese AI labs released major open-weight models in twelve days. Zhipu AI’s GLM-5.1 opened the sequence. MiniMax M2.7 followed. Then Moonshot’s Kimi K2.6, and finally DeepSeek V4 — both a Pro and a Flash variant. The pace was not incidental. Each lab watched the prior release and responded faster.

The models are not incremental. DeepSeek V4 Pro runs 1.6 trillion total parameters with 49 billion active, operates on Huawei Ascend hardware, and prices at under $2 per million tokens. On the Artificial Analysis Rails coding benchmark it scores 89/100. Kimi K2.6 — a 1T MoE model with 32B active parameters and a 256K context window — scores 87/100 on the same benchmark and costs approximately $4.50 per million tokens. Both are now listed among the leading open-weight models on the Artificial Analysis Intelligence Index, not just within their price tier.

GLM-5.1 ships under the MIT license. A 744B-parameter dense flagship and a 40B-active expert variant, both freely available for commercial use with zero royalty requirements. Zhipu’s stock closed 15.92% higher on launch day.

The pricing math is the story. Western frontier model APIs have held pricing disciplines that fund ongoing compute and research. This twelve-day window breaks that structure. Claude Opus costs over three times what DeepSeek V4 Pro charges per token. That gap is not a rounding error — it changes procurement decisions at enterprise scale.

DeepSeek V4 Pro running on Huawei Ascend hardware deserves extra attention. US chip export controls were assumed to be a durable constraint on Chinese AI performance. V4 Pro’s benchmark scores and pricing demonstrate that Ascend-based inference is now production-competitive on frontier-class MoE models. The constraint has been engineered around.

For developers doing cost sensitivity analysis right now: Kimi K2.6’s 256K context window at $4.50/M tokens beats the price/context ratio of several well-regarded Western alternatives. DeepSeek V4 Flash reduces that floor further for high-volume batch jobs. Both are open-weight — self-hostable if the security posture demands it.

The practical objections to Chinese open-weight models — data residency, supply chain trust, geopolitical exposure — are real and have not gone away. But they are now objections made against models that match or exceed Western frontier performance at a fraction of the cost. That is a different conversation than the one developers were having six months ago.

Western labs have responded with new model releases, not price cuts. That asymmetry is unsustainable at the rate Chinese labs are shipping.

The cost war is not approaching. It arrived in April, and the next wave is already being tested in private repos.