China's Open-Source AI Ecosystem: A New Era of Architectural Innovation

Architectural Choices in China’s Open-Source AI Ecosystem

The “DeepSeek Moment” of January 2025 marked a significant turning point in China’s open-source AI community, with a shift in focus from model performance to system design and architectural innovation. Chinese companies such as Huawei, Baidu, and Alibaba have been at the forefront of this movement, with notable achievements including the development of Mixture-of-Experts (MoE) architectures and the adoption of domestic hardware.

Why This Matters

The move towards MoE architectures and domestic hardware adoption has significant implications for the future of AI development in China. By prioritizing sustainability, flexibility, and cost-effectiveness, Chinese companies are able to develop AI systems that are better suited to real-world applications and constraints. However, this shift also poses challenges, such as the need for more permissive open-source licenses and the potential for increased competition in the market. The failure to adopt these new architectures and technologies could result in a significant loss of market share, with estimates suggesting a potential loss of up to 30% of the Chinese AI market.

Key Insights

• 20% reduction in training costs achieved by Ant Group’s Ling open models using optimized training on domestic AI chips, 2025
• Mixture-of-Experts (MoE) architectures have become the default choice for leading models from the Chinese community, including Kimi K2, MiniMax M2, and Qwen3
• Moonshot AI’s serving system, Mooncake, has been open-sourced, supporting features such as prefill/decoding separation and raising the baseline for deployment and operations across the community

Working Example

# Example of a simple Mixture-of-Experts (MoE) architecture
import torch
import torch.nn as nn

class MoE(nn.Module):
    def __init__(self, num_experts, input_dim, output_dim):
        super(MoE, self).__init__()
        self.num_experts = num_experts
        self.input_dim = input_dim
        self.output_dim = output_dim
        self.experts = nn.ModuleList([nn.Linear(input_dim, output_dim) for _ in range(num_experts)])
        self.gate = nn.Linear(input_dim, num_experts)

    def forward(self, x):
        gate_outputs = self.gate(x)
        gate_outputs = torch.softmax(gate_outputs, dim=1)
        expert_outputs = []
        for i in range(self.num_experts):
            expert_output = self.experts[i](x)
            expert_outputs.append(expert_output)
        expert_outputs = torch.stack(expert_outputs, dim=1)
        outputs = torch.sum(gate_outputs.unsqueeze(2) * expert_outputs, dim=1)
        return outputs

Practical Applications

• Use Case: Huawei’s Ascend AI chips have been used to achieve day-zero support for DeepSeek-V3.2-Exp, enabling developers to validate real-world performance directly.
• Pitfall: The use of prescriptive and tailored licenses can add friction to the adoption of open-source models, contributing to the decline of their usage.

References:

• https://huggingface.co/blog/huggingface/one-year-since-the-deepseek-moment-blog-2
• https://huggingface.co/spaces/cfahlgren1/hub-model-tree-stats
• https://www.huawei.com/en/news/2025/1/deepseek-v3-2-exp-ascend-ai-chips