5 Chinese models you need to test

In our first article of the year, I want to talk about a topic that has been gaining momentum: the advancement of open language models developed in China.

When we talk about large language models (LLMs), the debate usually revolves around OpenAI, Anthropic, Google, or Meta.

But that no longer reflects the real landscape.

China is reshaping the artificial intelligence landscape with a strong ecosystem of open source (or open weight) models, offering competitive performance, a focus on efficiency, and, in many cases, permissive licenses for commercial use.

In 2025, the narrative has changed: Chinese LLMs have moved from being promising alternatives to becoming leaders across multiple benchmarks.

The result?

More options for those who develop, research, or deploy GenAI in production, without relying exclusively on closed models.

In today’s article, we’ll look at the main open Chinese AI models that are truly worth experimenting with, focusing on recent models, open weights, and practical relevance.

If you have a spare GPU lying around or want to test new APIs, these are the names you need to know!

Follow our page on LinkedIn for more content like this! 😉

1. Qwen3 (Alibaba Cloud)

Developer: Alibaba Cloud
License: Apache 2.0 (most models)

The Qwen3 family is one of the most complete and versatile in the Chinese ecosystem. Developed by Alibaba, it is not a single model but a full family, with variants for text, code, multimodality, and long context.

Technical highlights:

• Parameters: versions of ~0.6B, 1.7B, 4B, 8B, 14B, 32B, and above

• Context window: 32k, 64k, or more in recent versions

• Strong performance in reasoning, code, and multilingual tasks

• Trained on 100+ languages and dialects (with support for Portuguese)

• Qwen3 models support “thinking mode,” which can be enabled or disabled

Why test it:

Qwen is extremely versatile. It works well both in local setups and in production and has excellent integration with agent tools and RAG.

Where to find it: Hugging Face – Qwen | GitHub

✨ See also: Multimodal RAG in Practice with Open-Source Models

---

2. DeepSeek (DeepSeek AI)

Developer: DeepSeek AI
License: MIT License

DeepSeek quickly became a reference when it comes to reasoning, mathematics, and code, especially in open source environments.

The model directly rivals proprietary models in complex reasoning and coding tasks, but with a training cost 30× lower.

Technical highlights:

• Parameters: ~1.5B, 7B, 8B, 14B, 32B, 70B, and above (Mixture-of-Experts variants)

• Context window: 128K tokens

• MoE architecture: DeepSeek-V3 has a total of 671B parameters, but activates only 37B per token

• Excellent performance in logic tasks, code generation, and complex problem solving

Why test it:

If you work with software engineering, agents, or automation, DeepSeek is a great choice.

Where to find it: GitHub – DeepSeek-V3 | DeepSeek-R1

✨ See also: Build Your Own Search System with EmbeddingGemma, SQLite, and Ollama

---

3. Kimi K2 (Moonshot AI)

Developer: Moonshot AI
License: Modified MIT License

Moonshot AI changed the game in November 2025 with the launch of Kimi K2, which outperformed proprietary models with a score of 44.9% on Humanity’s Last Exam.

Technical highlights:

• Context window: 128K tokens natively, expandable to 256K

• MoE architecture: 1 trillion total parameters, activating 32B per inference

• K2 Thinking: A reasoning variant that thinks while using tools, combining CoT with agentic execution

• MuonClip Optimizer: Moonshot’s proprietary optimizer that improves computational efficiency by 2× compared to AdamW

Why test it:

The model achieved state-of-the-art performance in several tasks involving knowledge, mathematics, and programming, and supports autonomous agents with native tool use.

Where to find it: GitHub – Kimi-K2 | Hugging Face

✨ See also: 9 Language Models with the Best Cost-Benefit for Running Locally

---

4. GLM 4 / ChatGLM (Zhipu AI / Z.ai)

Developer: Zhipu AI (now Z.ai)
License: Varies (Apache 2.0 for some)

The GLM family, developed by Zhipu AI (known as Z.ai), uses a unique architecture based on the “General Language Model” (GLM) framework, which differentiates it from others.

The most recent model, GLM 4.7, with 358B total parameters and 32B activated, quickly positioned itself as a new capability leader in the open ecosystem.

Technical highlights:

• GLM architecture: Instead of traditional GPT-style autoregressive modeling, it uses autoregressive gap filling, optimizing both understanding and generation

• Context window: 128K tokens in the open version

• Good balance between performance and computational efficiency

Why test it:

It is a stable, well-documented model and is increasingly used in corporate and academic applications. It shows gains in coding, interface and web page development, supports tool use, and has strong mathematical and reasoning capabilities.

Where to find it: GitHub – GLM-4 | Hugging Face

✨ See also: Practical guide to running Gemma 3 270M on your phone

---

5. MiniMax-M2 (MiniMax AI)

Developer: MiniMax AI
License: MIT License / Apache 2.0

MiniMax exploded onto the scene in 2025 with radical innovations in long context and computational efficiency.

The MiniMax-M2 model is a compact, fast, and cost-efficient MoE model (230 billion total parameters with 10 billion active parameters) built for coding and agent tasks.

Technical highlights:

• Record-breaking context: MiniMax-01 processes up to 4 million tokens, thanks to the Lightning Attention architecture

• Specialized in code and agentic tasks, standing out among open-source LLMs, especially for agentic tool calls

Why test it:

If your use case involves conversation, virtual assistants, or storytelling, it is well worth experimenting with.

Where to find it: GitHub – MiniMax-M2 | Hugging Face

✨ See also: LangChain with MCP: Connecting tools with flexibility and interoperability

Conclusion

The rise of these Chinese models represents more than just technical alternatives—it signals a fundamental shift in the AI ecosystem:

• True democratization: With MIT and Apache 2.0 licenses, any developer can download, modify, and even commercialize these models

• Radical cost reduction: Training state-of-the-art models for $5M instead of $100M completely changes the economics of AI

• Competitive benchmarks: They are no longer “almost as good”—in many cases, they are better

For developers, researchers, and companies, it’s worth evaluating these alternatives.

Many organizations are already building on top of these Chinese models, not out of ideology, but pragmatism: they work, they’re affordable, and they’re open.

How to Get Started

Most of these models are available via:

• Hugging Face for direct weight downloads

• Official company APIs (sometimes free or very inexpensive)

• Local frameworks like Ollama and LM Studio for running locally

• Cloud setups using vLLM, SGLang, or TensorRT-LLM

For those looking to reduce dependence on closed models, explore high-quality open-weight alternatives, or build more controllable solutions, these models are absolutely worth testing.

And you, have you used any of these models yet? Which one is your favorite?

Leave a comment! 💬