Alibaba released the Qwen4 base model family on November 16 and put the largest variant, Qwen4-Max, on its Qianwen platform the same day. The numbers that matter: 2.4 trillion total parameters, 71 billion activated per token, native vision, and a context window in the million-token range.

Sparse attention is doing the work

Qwen4-Max is a mixture-of-experts model built on Qwen Sparse Attention (QSA). The total parameter count is the headline; the 71 billion active per token is what decides the bill. Every request touches roughly three percent of the weights, which is why a model this size can be priced like a mid-sized one.

The architecture also explains the context window. Sparse attention is what keeps a million tokens from turning into a quadratic bill, and the same mechanism is why cache hits can be sold at ¥0.5 per million tokens.

Where it actually ranks

Two results are worth quoting because they are third-party. In CodeArena, a global blind test focused on front-end programming, Qwen4-Max placed sixth — the best result a Chinese model has recorded there. On Work360, an all-scenario modelling benchmark, it scored 52.69, more than ten percent above the previous generation.

Neither number makes it the best model available. Both make it competitive in the tier that most enterprise work actually runs in.

Two deployments with real numbers

Alibaba put two customer cases on the record. In automotive manufacturing, programming a single workpiece used to take 475 minutes of human time; with Qwen4-Max it takes 24. That is not a benchmark, it is a workflow being replaced.

The second case is more interesting for anyone building agents. A financial research deployment calls 90 MCP tools and returns answers with dual-source traceability — every claim can be walked back to two independent inputs. Tool calling at that scale is exactly where long context and cheap cache hits compound.

The price, and what it changes

Cache hits are ¥0.5 per million tokens. Alibaba says average cost per completed task is down about 60% against the previous generation. For chat, that is a modest saving. For agents, it is the whole argument: an agent loop resends the same system prompt and tool schemas dozens of times per task, so the cache-hit price is the real unit price, not the headline input price.

If your workload is a long-running agent with a stable prefix, the effective discount is far larger than 60%. If your workload is one-shot generation with a fresh prompt every time, it is smaller.

What is not settled

The smaller Qwen4 models will be open-sourced, but no date has been given, and open weights are the part of this release that changes the self-hosting calculus. There is also no independent evaluation of how quality holds up at the far end of a million-token window — vendors rarely publish that curve, and it is usually where long-context claims soften.

What to do with this

Do not switch on the leaderboard position. Run three of your own tasks: one short, one with a 200K-token document, one agent loop with a fixed prefix. Measure cost per completed task and cache hit ratio, not tokens per second. If the cache hit ratio on the agent task is high, the 60% claim will look conservative; if it is low, the model is simply a cheaper mid-tier option and nothing more.