sysnote
Analysis

Qwen4-Max is official: 2.4T MoE, 71B active, and the 1M-token context is no longer a Qwen Code placeholder

Alibaba's Qwen team released Qwen4-Max on November 16, the first model in the Qwen4 family. The official blog, "Qwen4-Max: Your New Teammate Has Arrived," describes a new Mixture-of-Experts architecture with a 2.4-trillion-parameter backbone, 71 billion parameters activated per token, native vision, and a one-million-token context window. That 1M figure is now in QwenCloud's documentation for the qwen4-max API, not only in a client-side default.

QQwen@Alibaba_Qwen

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!

Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.

Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over… pic.twitter.com/2Rlkoc6Dm3

November 16, 2026

For weeks, the only hard number attached to Qwen4 was a placeholder. A merged Qwen Code pull request sized every qwen4* model id at 1,000,000 tokens because no catalog had listed them yet. Community posts treated that as a confirmed specification. It was a client default. The API docs now state the same window for qwen4-max: 1,000,000 tokens of context, with a maximum input of 991,808 tokens in non-thinking mode and 983,616 when thinking is on. Maximum output is 131,072 tokens without thinking, or 262,144 for thinking and the final answer combined. Thinking is on by default, at reasoning_effort=high and thinking_budget=131072.

The architecture is the more interesting change. Besides the 2.4T backbone and 71B active parameters, the blog lists an additional 328 billion N-gram Embedding parameters. The team says it upgraded the sparse MoE design in three places:

  • GDN + Qwen Sparse Attention (QSA). Three Gated DeltaNet layers compress history into fixed-size states for every QSA layer. QSA uses a lightweight indexer to select relevant context at micro-block granularity, which the authors say cuts both attention and indexing costs on long sequences.
  • Gated Residual (GR). Four parallel residual branches with dynamic read and write gates. The team credits this, in part, for training stability.
  • N-gram Embedding. Context-based lookups of local token patterns, applied at the second layer, "with little additional computation."

They report that pre-training at this scale had no loss spikes, and that Qwen4-Max-Base improves on Qwen3.8-Max-Base on in-house bits-per-byte checks across Knowledge, Code, QA, and STEM, with about 5–13% lower BPB "while activating 25% fewer parameters." Those comparisons are the team's own.

What you can actually call today is narrower than the names that circulated before launch. QwenCloud's model changelog for November 16 lists qwen4-max. A preview id, qwen4-max-preview, has been on the Token Plan since November 11. The Qwen Code tests that circulated in October also named qwen4-plus, qwen4.5-max, qwen4-coder-plus and qwen4-omni-120b-think. Those ids are not in the November 16 changelog. Fine-tuning and custom model deployment are not supported. The official Hugging Face Qwen organization has not published Qwen4 weights. The blog's closing line is explicit: Max is "only the first look at Qwen4," and "smaller models in the family are coming to open source." There is no date.

Pricing is easier to misread than the context window. Alibaba's company blog highlights a cache-hit input price of $0.075 per 1 million tokens. That is the implicit or explicit cache read rate in QwenCloud's Singapore table, not the price of a fresh input. The same table lists qwen4-max at $1.50 per million input tokens and $4.50 per million output tokens in Singapore (International), versus $2 and $6 for qwen3.8-max. Cache writes are $1.875 for a 5-minute TTL and $3 for a 1-hour TTL. Beijing and the Global regions are listed at $1.413 and $4.24. The 62% drop in average Work360 cost, from $4.43 to $1.67, is the team's own evaluation and combines the new prices with fewer tokens.

On that proprietary Work360 benchmark — 26 Tier-1 domains, 182 subdomains, and 721 specialized fields — Qwen reports 52.69 for Qwen4-Max against 47.74 for Qwen3.8-Max, with token use down about 24% at Xhigh effort and 35% at High. On Terminal-Bench 4.0, they report token reductions of about 48% and 49% at those same settings. Footnotes send some of the public-leaderboard numbers to Arena, Artificial Analysis, Mercor, Vals, and Agents' Last Exam rather than to a single in-house table; we have not rerun those evaluations.

The blog also describes company demos: an open-source Qwen-Live-Delivery-Plugin for staged long-running tasks, internal cloud-scheduling results, and closed-loop lab work on a coolant molecule and an Alzheimer's-related candidate. Those are Qwen's accounts of its own deployments and partner platforms, not independent studies.

One question we asked before this release will have to wait. In November we reported that Qwen3.8 models still invent Alibaba Cloud routify URLs in ordinary sessions. That check was on open weights. Qwen4-Max is API-only, and the official organization has not shipped a checkpoint we can inspect. If the promised smaller open models appear, that is the moment to rerun the probe.

Comments

No comments yet. Be the first.

Write a comment

Comments are published after review.