guide
A structured snapshot of Chinese model families that are relevant for global evaluation in 2026: DeepSeek, Qwen, GLM, Kimi, ERNIE, Hunyuan, Pangu, Spark and Step.
DeepSeek and Qwen cover the main global developer baseline, GLM covers coding-agent evaluation, Kimi covers very long context and ERNIE/Qianfan cover Baidu Cloud or China-local enterprise deployment.
Based on the supplied May 15, 2026 China AI Navigator compilation, weighted toward model positioning, developer access, global usability and decision value.
The candidates serve different adoption questions, so the shortlist should not be treated as a single leaderboard.
Positioned in the supplied compilation as the open-source and coding baseline. Verify actual model availability and pricing before production use.
Better framed as the broad multimodal and multilingual model family for teams that care about Alibaba Cloud and open-model momentum.
Most relevant when the question is coding-agent depth, autonomous development workflows or China-market enterprise model options.
Use it as the long-context candidate for document-heavy workflows, research analysis, legal review and agent task execution.
Evaluate it first when Baidu Cloud, Chinese compliance or local enterprise deployment is already part of the stack.
Add StepFun when the evaluation includes multimodal open-source models, agent-focused Step 3.5 Flash or AI plus device commercialization.
Use Hunyuan when OpenAI-compatible API access, card-capable Tencent Cloud billing and Tencent Cloud ecosystem fit are central to the decision.
Evaluate Pangu for industry models, Ascend/Huawei Cloud infrastructure and enterprise deployment rather than consumer chat.
Add Spark when the workload is speech recognition, speech synthesis, education, healthcare or voice-heavy multimodal interaction.
Global users should start from workload, not brand. The same team may need one model for code, one for long documents and one for cloud-compliant enterprise rollout.
Start with DeepSeek, then add GLM when long-running coding-agent behavior matters.
Start with Qwen when language breadth, vision-language workflows or Alibaba Cloud deployment are central.
Use Kimi as the long-context candidate and compare against Qwen or DeepSeek for cost and API fit.