Alibaba, DeepSeek push China’s AI model race towards lower costs

Details

Alibaba's Qwen3.8-Max is its largest model to date, using a mixture-of-experts design that activates roughly 95 billion of its 2.4 trillion parameters per request. It supports 1 million token context, handles text, images, and video, and costs $2 per million input tokens and $6 per million output tokens. On Arena.AI's crowdsourced leaderboard it reached the top spot among Chinese text models, though it trails several Anthropic models in the overall ranking. It ranked second on the visual leaderboard, behind a Claude Fable 5 variant. Alibaba also reported the model completed a software engineering project over 16 continuous days. Moonshot AI's Kimi K3 (2.8T parameters, 104B active) is Qwen3.8-Max's direct scale competitor, priced higher at $3/$15 per million tokens. DeepSeek's V4-Flash takes a different position: 284B total parameters with only 13B active during inference, priced at $0.14/$0.28 per million tokens. Cache-hit pricing for the Max Effort version drops to $0.003 per million tokens — 98% below the standard input rate. Artificial Analysis estimated V4-Flash's average cost at $0.03 per benchmark test. The pattern shows Chinese AI providers now competing along two distinct axes: frontier scale (Alibaba vs Moonshot) and aggressive cost compression (DeepSeek). Both strategies pressure US providers on pricing, and the MoE architecture choices reflect lessons learned from DeepSeek's earlier efficiency-focused releases.