Alibaba spent three weeks staging Qwen4-Max: the family was announced at the Apsara Conference on 29 October, a subscription-only preview (qwen4-max-preview) opened on 11 November, and on 15–16 November the Qwen team finally published the production model. qwen4-max is now pay-as-you-go for anyone with a Singapore-region key — 2.4 trillion total parameters, 71 billion active per token, a one-million-token context, native text, image and video input, and a flat $1.50 input / $4.50 output per million tokens.
The architecture is finally disclosed, which it never was for Qwen3.8-Max. It is a sparse mixture-of-experts model mixing Gated DeltaNet 3:1 with Qwen Sparse Attention (QSA), plus 328 billion N-gram embedding parameters. Maximum output is 131,072 tokens without thinking and 262,144 with thinking plus answer — and some write-ups still quote 393,216, which was the preview’s ceiling, not the shipped one.
The pricing is the real story. The rate is flat across the entire 1M-token window with no tiering, and thinking mode costs exactly the same as non-thinking. Cache reads go for $0.075 per million, a five-minute cache write for $1.875, a one-hour write for $3.00, and new accounts get one million free tokens valid for roughly 90 days. Alibaba’s own number: an average Work360 agent task fell from $4.43 on Qwen3.8-Max to $1.67 on the new model, a mix of fewer tokens and a lower per-token rate. The international tariff is about 25% below Qwen3.8-Max’s.
Where it actually lands. On the Artificial Analysis Intelligence Index v5.0.1 it scores 42.9 (#15 of 58). Terminal-Bench 4.0 comes in at 48.5, DeepSWE v1.1 at 75.7, Vibe Code Bench v1.1 at 86.7, APEX-Agents at 72.1 and AutomationBench-AA at 72.5. It tops BioMysteryBench with 85.0 and holds the top spot in Code Arena WebDev at 1750 Elo. Head-to-head win-loss records read as a generational win but not a frontier one: 14–1 against its own Qwen3.8-Max-0902, 10–4–1 against Kimi K3.1 Pro, but 4–11 against Claude Opus 5.5 and 3–11 against Claude Sonnet 5.5.
Two caveats worth keeping: the weights are closed, and no Qwen4 model appeared on Hugging Face at launch — smaller ones are promised later with no date. Batch API is not supported on the international model page, the preview retires on 25 November at 16:00 UTC, and fine-tuning is unavailable.
Sources: Qwen blog · Artificial Analysis · llmlearner model card · launch guide and benchmarks