← November 16, 2026 briefing · added in the 20:00 KST update

Qwen4-Max follow-up: 2.4T total parameters, 25% fewer active than predecessor; external evals show ~67% lower cost per task (Korean media)

Announced November 16, 2026

Verified (fact)

According to AI Times, Alibaba's Qwen team published Qwen4-Max's architecture, benchmarks, and API pricing on its official blog on the 16th. This came five days after the preview opened on the 11th, when none of this information was disclosed. Total parameters are about 2.4 trillion, the same as the predecessor. However, the number of parameters actually used in computation was cut by 25%. Reports say external evaluations showed improved performance and roughly 67% lower cost per task. AI Newspaper (인공지능신문) reported that Alibaba plans to extend the model to agentic tasks for industrial and scientific research, such as manufacturing process automation and drug-candidate molecule design. Neither report included specific API prices or benchmark figures.

Needs fact-check (claim)

Alibaba claimed it reduced the number of tokens needed for reasoning and the cost per task, and improved the model's ability to handle complex work using multiple tools (as relayed by reports).

Why it matters (interpretation)

This is the first disclosure of the scale and architecture details that were missing at the preview. It is a case of lowering cost by reducing active parameters in a 2T-class model, giving enterprises a reference point for price-performance comparisons when choosing models. An earlier briefing reported the cost reduction as '62%'. Because this differs from the 67% in the latest reports, the measurement conditions should be checked in the original blog post.

Confidence medium · official source pending

Sources

Story thread

  1. 2026-10-29 "Qwen4 models support 1M-token context" spotted in qwen-code repo PR; Qwen4-max date appears to be a placeholder (community report, unconfirmed)
  2. 2026-10-29 Qwen4 follow-up: "Alibaba Cloud CTO said at Apsara Conference it will launch next week" (community report, unconfirmed)
  3. 2026-11-03 Qwen4 follow-up: qwen-code repo PRs list model IDs including qwen4-max, qwen4-coder-plus and qwen4-omni-120b-think (community report, unconfirmed)
  4. 2026-11-06 Qwen 4 follow-up: official Qwen X post previews Qwen 4 sizes, says scale will grow in Qwen 4.5 and 5 (community reports, unconfirmed)
  5. 2026-11-11 Qwen4 follow-up: 'Qwen4-Max Preview' released, up to 90% off on Alibaba Token Plan for a limited time (community report)
  6. 2026-11-12 Qwen4-Max Preview follow-up: Alibaba Token Plan only until Nov 25, 1M-token context; benchmarks and weights undisclosed (report)
  7. 2026-11-16 Alibaba fully launches Qwen4-Max after preview, claims 62% lower cost per task and filtered-out offensive cyber skills (reports, community posts)
  8. 2026-11-16 Qwen4-Max follow-up: 2.4T total parameters, 25% fewer active than predecessor; external evals show ~67% lower cost per task (Korean media)

More about Alibaba