Alibaba shipped Qwen4-Max this morning. New flagship, generally available, pay-as-you-go.
The headline numbers: 2.4 trillion total parameters on a new MoE architecture, with only 71B activated per token. Native vision. One million tokens of context.
On benchmarks, the official figures put Qwen4 sixth on CodeArena's global front-end coding blind test, and in the leading group on agentic evals like ALE and Vals Index. On Terminal-Bench 4.0, at the same Xhigh thinking setting, average tokens per attempt dropped from 289K to 151K while the score rose from 38.9 to 48.5.
Those are the vendor's numbers. Independent evals haven't caught up yet. Worth remembering before anyone declares a winner.
Price is the more interesting part. Cache hits run as low as 0.5 yuan per million tokens, and Alibaba says the average cost per task is down 60% against the previous generation. International pricing is $1.50 input / $4.50 output per million tokens.
For people who actually run models all day, that matters more than a benchmark table. Benchmarks are for launch day. The bill is for every day after. A 60% cut changes which tasks are worth running at high thinking settings, and it makes long-horizon agent work and batch jobs materially cheaper.
How to try it: the API is live on the Qwen AI platform, it's also available through the Token Plan, and it's built into Qwen Office, Qoder, and the Qwen app. Smaller models in the Qwen4 family are promised as open source later; Max itself is hosted only.
A quick timeline: a preview ran on the Token Plan from November 11. Today is the real general availability. The preview retires on November 25.
Zoom out a little. The Chinese model scene has been busy lately: DeepSeek opened its app features to API developers, then news broke of usage limits on its consumer app. Now Alibaba answers with a new flagship at a much lower price.
Good news for users, either way. More models, lower prices. Pick based on your own workload, not the launch charts. I haven't run my own tests on Qwen4-Max yet. When I do, I'll write that up separately.
Top comments (0)