Qwen4-Max Is Officially Out: 2.4T Parameters, AA Index 43
On November 16, 2026, Alibaba officially released Qwen4-Max, the flagship of the Qwen4 family. Only five days after the limited preview opened on November 11, this launch fills in the three things the preview was missing: parameter counts, benchmark scores and API pricing. The full release is already live on QwenCloud, Qwen Chat, Qwen App, Qwen Studio, QwenWork and Qoder — the regular entry points now run the full version, not a limited trial. The preview model ID, qwen4-max-preview, still runs until 24:00 Beijing time on November 25 as originally planned before retiring; it has simply been replaced early by the full release.
What the full release adds
- Parameters: a new Qwen4 architecture with 2.4 trillion total parameters and 71 billion active per token, plus a 328-billion-parameter N-gram Embedding module; native vision and a 1 million token context. Maximum output actually shrank versus the preview: 131,072 tokens without thinking, and 262,144 tokens shared between thinking and answer with thinking on (the preview shared 393,216).
- Scores, in two layers: Alibaba's own Work360 industry-task score is 52.69, about 5 points above the previous Qwen3.8-Max at 47.74, and on agent benchmarks such as DeepSWE it claims to beat Claude Opus 5.5 and GPT-6.1 Sol — that is the vendor's own claim, not independently verified. The independent Artificial Analysis Intelligence Index (xhigh reasoning effort) gives it 43: 9 points above Qwen3.8-Max's 34, 3 below Kimi K3.1 Pro's 46 and 16 below Claude Opus 5.5's 59.
- Pricing: pay-as-you-go API at $1.50 per million input tokens and $4.50 per million output tokens, with cached input billed at 5% (95% cheaper); developers can also stay on the Token Plan subscription. On cost, Alibaba's Work360 figures put the average cost per task down from $4.43 to $1.67, a 62% drop; on Artificial Analysis's measure, cost per task fell from $10.02 to $3.27.
- Access: regular users just pick Qwen4-Max in Qwen Chat, Qwen App, Qwen Studio, QwenWork or Qoder — no subscription needed; developers go through QwenCloud's Token Plan or the API.
What is still missing
- No open weights: the Max line keeps its old approach — cloud only, no downloads. Alibaba says smaller models in the family will be open-sourced later, with no size, license or date given.
- No independent latency data: Artificial Analysis has not published per-task timing yet, so how fast it actually feels is still unverified.
- The full release says nothing about dates for the Plus and Flash versions, and the 27B open-source version is dateless too.
Preview recap: what happened in the two weeks from November 11
When the preview opened at 10:00 Beijing time on November 11, it offered only a 1 million token context, multimodal input and free entry points — no parameters, no scores, no pricing — and was set to retire at 24:00 on November 25. The full release arrived 9 days early to take over; the preview ID itself still runs until 24:00 on November 25, but the regular entry points have already switched to the full version, so there is no deadline left to chase if you want to try it.
Also that day: Qwen Code 1.0
On November 11, Alibaba also released Qwen Code 1.0, merging the earlier Qoder and Qwen Code Harness into one open-source command-line coding tool. The full Qwen4-Max is available in Qoder as well, and coding remains one of the most worthwhile directions to try in this release.
Sources
- Qwen official blog: Qwen4-Max: Your New Teammate Has Arrived
- Artificial Analysis: Qwen4-Max scores 43 on the Intelligence Index (confirmed via the @ArtificialAnlys post on 2026-11-16 and the model page)
- Cocoloop: Qwen4-Max is out, scoring 43 on an independent index
- Ifeng: 2.4 trillion parameters! Alibaba's Qwen4-Max flagship debuts, tasks 58% cheaper
- Earlier coverage on this site: /news/news-qwen4-max-preview/ (published 2026-11-13, the preview version)