Chinese AI Models Rise: Qwen 3.6, Kimi K2.6, and GLM-5 Coding Benchmarks Review
Chinese AI Models in 2026: A New Era for Coding
For the first time in AI history, Chinese-developed models are competing at the highest level for coding tasks. In May 2026, three models have broken into the global top 15 on SWE-bench Verified.
The Rankings
Kimi K2.6 (Moonshot AI) - 80.2%
The surprise leader of Chinese models. With a 1 trillion parameter MoE architecture, Kimi K2.6 scores 80.2% on SWE-bench Verified and an impressive 85% on LiveCodeBench.
- Architecture: 1T MoE (Mixture of Experts)
- SWE-bench Verified: 80.2%
- LiveCodeBench: 85%
- Price: $0.60/$2.50 per million tokens
- License: Open-weight
Kimi K2.6 excels at competitive programming and offers the best value among frontier models. At $0.60/M input tokens, it's 8x cheaper than Claude Opus.
Qwen3.6 Plus (Alibaba) - 78.8%
Alibaba's flagship model continues the Qwen series' tradition of open-weight excellence. The 27B dense variant scores 77.2% with full Apache 2.0 licensing.
- Architecture: Dense and MoE variants
- SWE-bench Verified: 78.8% (Plus), 77.2% (27B)
- LiveCodeBench: 83.6%
- Price: $0.50/$2 per million tokens
- License: Apache 2.0
Qwen3.6 is the best open-weight model you can self-host. The Apache 2.0 license means true freedom to modify and deploy.
GLM-5 (Zhipu AI) - 77.8%
Zhipu's 744B parameter model trained on domestic Huawei chips. The most "Chinese-developed" of the top models.
- Architecture: 744B parameters
- SWE-bench Verified: 77.8%
- LiveCodeBench: 52%
- Training: Huawei Ascend chips
- License: Open-source
GLM-5 shows that domestic chip training can produce world-class models. Its coding ability is impressive, though LiveCodeBench score lags.
How They Compare to US Models
| Metric | Claude Opus 4.7 | Kimi K2.6 | Qwen3.6 | GLM-5 |
|---|---|---|---|---|
| SWE-bench V | 87.6% | 80.2% | 78.8% | 77.8% |
| LiveCodeBench | - | 85% | 83.6% | 52% |
| Price/M tokens | $5/$25 | $0.60/$2.50 | $0.50/$2 | - |
Chinese models are 7.5-10x cheaper than Claude Opus while delivering 90-92% of the coding performance.
Real-World Usage
In February 2026, Chinese open-source models accounted for over 50% of global token consumption for the first time:
- Moonshot AI (Kimi): 14.5%
- DeepSeek: 9.0%
- MiniMax: 4.2%
The Verdict
Chinese AI models are no longer "catching up" - they're competing at the highest level. For developers who need:
- Best Chinese model: Kimi K2.6 (80.2% SWE-bench)
- Best open-weight: Qwen3.6 (Apache 2.0)
- Most domestic: GLM-5 (Huawei chip trained)
The price advantage is enormous. If your workflow can tolerate slightly lower benchmark scores, these models offer incredible value.