Alibaba Qwen3.8-Max Explained 2026: 2.4T MoE, 1M Context, 86.6 Terminal-Bench

The China AI cascade just got another entry. On August 4 Alibaba unveiled its biggest model yet, Qwen3.8-Max — a staggering 2.4 trillion parameters, though it only activates 95 billion per query as a mixture-of-experts model. On benchmarks Alibaba claims it approaches or edges Anthropic’s Fable 5 and Claude Opus 5. Here’s the gist.

What makes it notable?

• Architecture: sparse MoE, 95B active of 2.4T total parameters
• Context: 1M tokens — 200+ pages of text or ~100 hours of footage per request
• Multimodal: text, image and video input
• Alibaba’s largest and most capable Qwen to date

It’s the peak of the “build it huge, activate only a slice” trend. The same China open-model wave continues in DeepSeek V4 Flash.

How far do the benchmarks go?

Per Alibaba’s published table:

Item	Qwen3.8-Max	Note
Frontend Code Arena	1,668	37 points behind Opus 5’s top config
PaperBench	93.0	Leads
IFBench	82.8	Leads
Terminal-Bench 2.1	86.6	Ahead of Opus 4.8 & Fable 5 (84.6)

The point is it stands shoulder to shoulder with the top tier on coding and hands-on metrics. China’s rise also shows in AI rankings via OpenRouter.

Pricing and availability?

• Pricing (per 1M tokens): $2 input / $6 output / $0.25 cached input
• Available now via QwenCloud; weights to be open-sourced on Hugging Face and ModelScope

Releasing top-tier performance as open weights piles real pressure on the industry.

Frequently Asked Questions

Q. 2.4T parameters — why is it called fast? It’s a mixture-of-experts model, so each query activates only about 95B parameters rather than the whole thing. That keeps a very large model relatively fast and cheap.

Q. Should I trust the benchmarks as-is? These are Alibaba’s own published numbers. The direction is clear, but independent benchmark verification is worth waiting for.

Q. If it’s open-weight, can I use it commercially? Weights are set to publish on Hugging Face and ModelScope. Check the exact license terms at release.