The current state of the Chinese LLMs scene

The Chinese large language model (LLM) scene is a crowded and competitive field with distinct leaders and innovators. ByteDance, the creator of TikTok, holds the market leadership position with its proprietary model Doubao, playing a role analogous to OpenAI in the West. However, the most technically innovative player is DeepSeek, a side project from an algorithmic trading firm, which is a close second in user adoption. DeepSeek has invented key architectural techniques like MLA (Multi-head Latent Attention), DSA, and GRPO that are being adopted by other Chinese AI firms.

Beyond the proprietary giants, the open-weight model space is fiercely contested. Alibaba's Qwen series is strongest in its open-weight offerings, particularly for smaller models, and leads in text-to-image (T2I) and text-to-video (T2V) generation. Meituan is emerging as an aggressive open-weight player with its 562-billion parameter LongCat-Flash-Chat model, which uses a dynamic mixture-of-experts (MoE) architecture activating 18.6B to 31.3B parameters per token. This 'open model race' is further fueled by the 'Six AI Small Tigers'—companies like Zhipu, Minimax, and Moonshot—which release large open models to gain recognition and provide low-cost inference services, though their long-term viability remains uncertain.