Alibaba released Qwen4-Max overnight, the new top model in its Qwen line. According to Bloomberg, it beats Claude Fable 5.1 and Moonshot’s Kimi K3.1 Pro on coding benchmarks. You can’t download this one. Alibaba says smaller Qwen4 models with open weights will follow soon.
What’s inside
Qwen4-Max is a mixture-of-experts model with a 2.4-trillion-parameter backbone, 71 billion of which are active per token. On top of that come 328 billion parameters of n-gram embeddings that look up common local token patterns. The context window holds a million tokens, and vision is built in. Alibaba says training ran without a single loss spike and credits the new architecture: Gated DeltaNet layers, its own sparse attention scheme called QSA, and four parallel residual branches.
Fewer tokens, lower prices
The biggest jump Alibaba claims is in efficiency. On Terminal-Bench 4.0, Qwen4-Max needs roughly half the tokens of its predecessor Qwen3.8-Max; on Alibaba’s own Work360 benchmark it uses a quarter to a third fewer. Alibaba also cut API prices, especially for cached input. Put together, the average cost per Work360 task drops from $4.43 to $1.67. The score itself rises from 47.74 to 52.69.
Work360 spans 721 specialist fields, from finance to law to medicine, and the examples in the blog post are just as broad. Qwen4-Max builds scenes in Blender and makes playable games in Unreal Engine and Godot. At a car plant, the effort to program a single workpiece fell from about 475 minutes to 24. For pre-IPO due diligence, the model sorted 198 messy working papers into 14 folders and wrote an 86-page report.
Qwen4-Max is available now on QwenCloud, via the API or a token plan.
Alibaba benchmarked the previous model
The comparison is against Claude Fable 5.1. But Anthropic shipped Claude Fable 5.2 on November 9, with the same context window and prices about 20 percent lower. How Qwen4-Max stacks up against the current model, the numbers don’t say. For a model built this obviously around long work tasks, that’s the question that matters.
The cost story is still worth watching. If you run a lot of agent steps, what you end up counting is dollars per finished task, and that’s exactly where Alibaba is pushing. I’m curious where it lands on the independent leaderboards once the first runs against Fable 5.2 come in.