China's Z.ai (Zhipu AI) and Moonshot AI released new-generation models, GLM-5.2 and Kimi K2.7-Code, in quick succession in June 2026. As the US restricts the sharing and export of frontier models, Chinese labs are foregrounding an open-weight strategy and competing on low cost.
June 2026 · China Open-Weight AI Race
Two Chinese Labs Ship Open Coding Models Days Apart
As the US tightens export controls on frontier models and chips, Z.ai's GLM-5.2 and Moonshot's Kimi K2.7-Code double down on an open-weight strategy — racing to close the gap with closed US models.
1M
GLM-5.2 context window (tokens)
+21.8%
Kimi K2.7 gain on Kimi Code Bench v2 vs K2.6
−30%
Reasoning-token usage cut by K2.7-Code
Context window, drawn to scale
GLM-5.2 jumps to ~4× the prior 200K, and roughly 4× Kimi's window — each block = ~100K tokens.
GLM-5.2
Z.ai (Zhipu AI)
Context: 1M tokensLicense: MIT (planned)Access: GLM Coding Plan / APIModes: High & Max thinking, async Agent RLHardware: trained on Huawei Ascend — zero NVIDIA reliance
Kimi K2.7-Code
Moonshot AI
Context: ~256K–262K tokensLicense: Modified MITAccess: Hugging Face / Kimi APIBenches: +11% Program Bench, +31.5% MLS Bench LiteFocus: agentic coding, leaner reasoning
What's working
GLM-5.2 praised for one-shot coding speed
Honest kernels on KernelBench-Hard — no reward hacking
Fast shipping, strong open-model catch-up, good cost & access
Caveats
Kimi K2.7 reported slow on the hosted side
Gains rest on internal benchmarks, limited independent verification
Fable 5 still seen as strong; latency & quota can bottleneck
The bigger picture
Open-weight Chinese models are closing in on closed US frontier systems — sharpening a "behind the wall vs. open" divide over developer access.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…