What Happened
In early May 2026, China's AI community witnessed an intense "open-source arms race." Within just 12 days, four Chinese AI labs successively released open-weight code models:
| Date | Lab | Model | Background |
|---|---|---|---|
| ~May 1 | Zhipu AI (Z.ai) | GLM-5.1 | GLM-4 had already been widely praised in the open-source community |
| ~May 3 | MiniMax | M2.7 | Known for its cost-effective speech and text models |
| ~May 4 | Moonshot AI | Kimi K2.6 | Part of the Kimi product line, focusing on long context windows |
| ~May 5 | DeepSeek | V4 | The latest iteration of the DeepSeek series, following the international success of V3 in the open-source community |
This is no coincidence. The fact that four labs released competing products within nearly the same time window indicates that China's AI industry is entering a period of concentrated breakthroughs.
Key Facts
Performance Matches Western Frontiers
According to aggregated data from AI Tools Recap, these four models have already matched the performance levels of Western frontier models in agentic engineering benchmarks. Specifically:
- Performance on code task benchmarks like SWE-Bench and Terminal-Bench is close to that of Claude Opus 4.7 and GPT-5.5
- Demonstrates capabilities comparable to closed-source models in agent scenarios such as code generation, debugging, and refactoring
Even More Striking Cost Advantages
Inference costs are no more than a third of Claude Opus 4.7.
What does this mean? If your company needs to run thousands of code inference tasks daily, the cost of using these four open models could be just 30% of what it would be with Claude or GPT. For budget-conscious SMEs and startups, this is a highly attractive option.
Why Now?
Maturation of Technical Accumulation
Chinese AI labs have accumulated substantial engineering experience over the past two years. DeepSeek V3 already drew international attention in late 2025, and Zhipu AI's GLM series has continuously improved in both Chinese and English bilingual capabilities. The release of these four models is a natural outcome of this technical accumulation.
Accelerating Effects of the Open-Source Strategy
Open weights mean anyone can download, fine-tune, and deploy these models. This strategy has two effects:
- Rapidly expanding the user base--The developer community is the best channel for product dissemination
- Building ecosystem moats--Once the toolchains and communities built around a model take shape, switching costs become extremely high
Geopolitical Context
Notably, when the Pentagon signed an AI cooperation agreement on May 5, Chinese models were clearly not under consideration. On May 8, the US government confirmed AI model review agreements with Google DeepMind, Microsoft, and xAI. These developments have further reinforced the path of "self-reliance" for China's AI industry.
Impact on the Industry
Open Models Are Eroding the Moats of Closed Models
If open models approach closed models in performance while costing only a third as much, where do closed models retain their advantage? The answer likely boils down to security, compliance, and ecosystem integration. However, for a vast number of small-to-mid-scale use cases, these advantages are insufficient to justify the massive price difference.
Chinese AI Shifts from "Follower" to "Contender"
The narrative of a "Chinese version of GPT" is no longer relevant. When four Chinese models collectively benchmark against Western frontiers within 12 days, China's AI industry has evolved from a "catch-up player" into a substantive competitor.
Continuous Decline in Inference Costs
The release of these four models further compresses the cost curve for AI inference. The lower the inference cost, the lower the barrier to deploying AI applications. Ultimately, all AI users benefit--whether developers, enterprises, or individuals.