Chinese LLM choices narrow to DeepSeek, Qwen, GLM and Kimi

China’s open-weight LLM field is framed around DeepSeek, Alibaba’s Qwen, Z.ai’s GLM and Moonshot’s Kimi, which together cover production choices from Qwen3.8-27B on one GPU to Kimi K3, a 2.8 trillion parameter model requiring a multi-node cluster. Sparse mixture-of-experts designs keep active parameters low per token, but total parameter count still determines hosting costs.

DeepSeek V4 Pro is presented as the strongest default flagship because it combines MIT weights, competitive independent ranking and lower pricing than other open flagships. DeepSeek V4.1 Flash and GLM 5.3 Flash lead the Flash-class shortlist: DeepSeek is described as newer and faster, while GLM scores slightly higher on the cited index and supports video input.

Qwen has the broadest lineup, including Qwen 3.8-Max and Qwen3.8-27B, the only model listed as practical on a single GPU. GLM is positioned for long-horizon agent and engineering sessions, with GLM 5.3 Flash and Qwen 3.8-Flash-Next called out as the cheapest API options at $0.15 per million input tokens and $0.50 or $0.47 output. Kimi K3 is the largest and most heavily benchmarked open flagship, but its size makes API access the practical route for most teams.

Two pending releases could change the ranking: Qwen 4, described by Alibaba as in training and coming “very soon,” and DeepSeek V4.1 Pro, confirmed but undated.