Four Chinese Labs Rewrote the Open-Weights Leaderboard in 18 Days

Between April 7 and April 24, four Chinese AI labs shipped open-weight coding models in close succession: Z.ai's GLM-5.1, MiniMax's M2.7, Moonshot's Kimi K2.6, and DeepSeek's V4. All four are competitive with Western frontier models on agentic engineering and coding benchmarks. All four run inference at under one-third of Claude Opus 4.7's pricing. And none of them were trained on Nvidia hardware.

That last detail is the one I keep returning to. The US export controls on H100 and H800 chips were intended to create a compute gap, and for a while, the conventional wisdom held that they had. The April releases suggest otherwise. GLM-5.1 was trained entirely on 100,000 Huawei Ascend 910B chips. DeepSeek V4 is a 1.6 trillion parameter model built with zero Nvidia hardware. The compute restriction pushed Chinese labs toward algorithmic efficiency as a substitute for raw scale, and what came out the other side is a cluster of models that are now within 3-6 Intelligence Index points of the Western frontier.

The benchmark picture is real, if messy. GLM-5.1 briefly became the first open-weight model ever to top SWE-Bench Pro when it dropped on April 7, scoring 58.4%. It held that spot for nine days until Claude Opus 4.7 reclaimed it on April 16. Kimi K2.6 hit 58.6% on SWE-Bench Pro on its own, edging out GPT-5.4. On Artificial Analysis's Intelligence Index, Kimi K2.6 and Xiaomi's MiMo V2.5 Pro tie at 54, with DeepSeek V4 at 52. For context, GPT-5.5 sits at 60, Opus 4.7 at 57. The gap is real but narrow.

What's harder to dismiss than the benchmark gap is the price gap. Running a million output tokens through Claude Opus 4.7's API costs $75. Running the same through Kimi K2.6 costs roughly $0.60. DeepSeek V4-Flash is $0.28 per million output tokens, which works out to about 0.4% of Opus pricing. For teams running coding agents at volume, this is not a marginal consideration. It changes the build economics entirely.

Each model made a different architectural bet. DeepSeek V4 introduced Engram conditional memory, an external O(1) knowledge lookup that feeds directly into the transformer backbone, separating memory from compute and making million-token contexts more efficient. Kimi K2.6 ships with an Agent Swarm primitive that runs a task across up to 300 sub-agents over 4,000 coordinated steps. MiniMax M2.7 describes itself as having participated in over 100 autonomous rounds of scaffold optimization during development, reporting a 30% gain from the loop. The "model trains itself" framing is overstated, humans still own the training run, but evaluator-in-the-loop scaffold tuning is real, and M2.7 is the most publicly committed bet on it.

The Anthropic and OpenAI teams did not sit still during this window. Claude Opus 4.7 shipped on April 16, right in the middle of the Chinese sprint, and reclaimed the SWE-Bench Pro crown. GPT-5.5 followed on April 23, one day before DeepSeek V4. The frontier moved while the open-weight wave was rising. The Western closed-source models still hold a genuine lead on hallucination rates and broad reasoning benchmarks, and the Elo gap on hard knowledge tasks remains meaningful.

But I think the framing of "caught up or not" misses the operational shift. The top 10 open-weight models on Artificial Analysis's Intelligence Index are now all from Chinese labs. Self-hosting a frontier-competitive coding model under an MIT license on Huawei hardware was not a viable option a year ago. It is now. GLM-5.1 in one documented demo ran 655 iterations with over 6,000 tool calls to build a vector database from scratch, running for up to 8 continuous hours without performance degradation. That is a different kind of machine than what open-source AI was even six months ago.

The 2026 LLM stack is no longer Anthropic vs. OpenAI vs. Google with everything else as a footnote. It's two regional capability pools at a 5–25x price gap, increasingly interoperable, with the cheaper pool shipping architecture papers that the expensive pool will be reading carefully.