Alibaba Unveils Qwen3.8-Max: 2.4 Trillion-Parameter MoE Model, China's Largest Yet

Summary: Alibaba's Qwen3.8-Max is a 2.4 trillion-parameter mixture-of-experts model that activates only 95 billion parameters at inference time, positioning it as China's most capable open-source AI to date.

Key Facts

• Scale: 2.4 trillion total parameters in a MoE architecture; 95 billion active per forward pass — largest in the Qwen family.
• Multimodal: Accepts text, image, and video inputs; outputs text. Covers the full multimodal stack in a single model.
• Benchmarks: Ranked #1 among Chinese models and #2 globally on the Arena.AI vision leaderboard (behind a Claude Fable 5 variant); trails leading Anthropic models on text tasks overall.
• Availability: Live now via Alibaba Cloud Model Studio APIs and QwenWork enterprise agent platform; open-source weights scheduled for public release next week.

Why It Matters

The gap between Chinese AI labs and US frontier labs is narrowing measurably. When Qwen3.8-Max's weights go public, developers worldwide will have access to a 2.4T-parameter MoE model at near-zero inference cost on their own hardware — a meaningful shift for cost-sensitive enterprise and research use cases. It also signals Alibaba's return to an open-source strategy after briefly moving flagship releases behind paywalls earlier this year.

Further Reading

• MarkTechPost analysis — MarkTechPost
• Bloomberg coverage — Bloomberg