Chinese AI Labs Dominate Open-Weight Models with Qwen, Moonshot, and Z.ai

---

This month has seen a significant shift in the landscape of open-weight models, with several impressive releases from Chinese AI labs. While I’ve always had a soft spot for Mistral, Gemma, and Llama, the latest models from Qwen, Moonshot, and Z.ai have set new benchmarks that are hard to ignore.

What’s New and Why It Matters

Over the course of July 2025, Chinese AI labs released a series of models that not only outperform their Western counterparts but also come with more accessible licensing. Here’s a breakdown of what was released:

• Moonshot Kimi-K2-Instruct (11th July): A massive 1 trillion parameter model.
• Qwen3-235B-A22B-Instruct-2507 (21st July): 235 billion parameters, designed for instruction following.
• Qwen3-Coder-480B-A35B-Instruct (22nd July): 480 billion parameters, optimized for coding tasks.
• Qwen3-235B-A22B-Thinking-2507 (25th July): Another 235 billion parameter model, this time focused on reasoning and problem-solving.
• Z.ai GLM-4.5 and GLM-4.5 Air (28th July): Two models with 355 billion and 106 billion parameters, respectively.
• Qwen3-30B-A3B-Instruct-2507 (29th July): A more manageable 30 billion parameter model for instruction following.
• Qwen3-30B-A3B-Thinking-2507 (30th July): Another 30 billion parameter model, this time for reasoning tasks.
• Qwen3-Coder-30B-A3B-Instruct (31st July): A 30 billion parameter coding-focused model.

Key Details and Implications

Licensing

One of the standout features of these models is their licensing. While Moonshot’s Kimi K2 uses a non-OSI-compliant modified MIT license, all Qwen models are licensed under Apache 2.0, and Z.ai’s models use the MIT license. This makes them more accessible for both commercial and academic use.

Accessibility

The larger models from Chinese labs offer their own APIs, making it easier to integrate them into existing workflows. Additionally, some of these models, like the Qwen 30B and GLM-4.5 Air 106B, can be run on personal laptops, which is a significant advantage for developers who want to experiment without heavy infrastructure.

Performance

The performance of these models is nothing short of impressive. For example, the Moonshot Kimi-K2-Instruct with its 1 trillion parameters sets a new standard in terms of scale and capability. The Qwen3-480B-A35B-Coder is particularly noteworthy for its coding capabilities, while the smaller 30 billion parameter models offer a balance between performance and resource efficiency.

Notable Absences

It’s worth noting that DeepSeek, another prominent player in the AI space, hasn’t released any new models since April. Their last release, DeepSeek-R1-0528, is available on Hugging Face but doesn’t match the recent advancements from Chinese labs.

OpenAI and the Competition

The delay in releasing OpenAI’s open-weight model has raised some eyebrows. It’s possible that they are aiming to launch a model that can stand out against this impressive lineup from Chinese AI labs. Indeed, when OpenAI did release their models on August 5th, they were highly competitive.

Conclusion

The rapid advancements and open-source nature of these Chinese AI models have significant implications for the AI community. They not only push the boundaries of what’s possible but also make cutting-edge technology more accessible to a broader audience. As we move forward, it will be exciting to see how these models are adopted and integrated into various applications.