subagentic.ai RSS X YouTube GitHub
AI-generated illustration

News

Qwen4-Max is generally available, with a 1M window and no open weights

Alibaba makes Qwen4-Max generally available: a 2.4T multimodal MoE with a 1M context window and published API prices.

qwenalibabamodelsagents

Alibaba released Qwen4-Max on November 16, 2026, from Hangzhou, and called it the most capable model in the Qwen family. The announcement describes a sparse mixture-of-experts model with 2.4 trillion total parameters, native vision, and a one-million-token context window. TechNode reported the same day that the model is generally available on QwenCloud, following a limited preview that began on November 11.

The announcement does not include open weights for Qwen4-Max. Alibaba said smaller open-weight models in the series will follow.

Size, window, and where to call it

Alibaba says Qwen4-Max activates 71 billion parameters per token, 25% fewer active parameters than Qwen3.8-Max at the same 2.4 trillion scale. It credits a Qwen4 architecture that upgrades the sparse MoE design across attention, residual connections, and embeddings, and says the smaller active set is meant to deliver stronger capability at a lower inference cost.

The model is available now on QwenCloud via the Token Plan and the API. Alibaba also lists QwenWork and QwenStudio. TechNode adds Qoder.

The only token rate in the press release is an implicit cache hit on input, priced at $0.075 per million tokens. Standard input and output list prices are not stated in that release or in the TechNode recap. TechNode says lower API prices, together with fewer tokens than Qwen3.8-Max on comparable tasks, brought average task cost on Alibaba’s Work360 benchmark from RMB 26.58 to RMB 11.16, a reduction of about 58%. Alibaba’s release puts that Work360 cost drop at around 60%, and says the model uses 35% fewer tokens than Qwen3.8-Max in High effort reasoning mode.

A model id string is not given in either source. Support for fine-tuning, dedicated deployment, or batch inference is also unstated.

No weights with this release

Alibaba said open-weight, smaller-scale models from the Qwen4 series will soon be released to developers worldwide. TechNode reports that Alibaba plans to release smaller Qwen4 models as open source. Neither write-up says weights for Qwen4-Max shipped with the November 16 release.

Benchmarks and showcase runs, as Alibaba tells them

Alibaba says Qwen4-Max ranked 6th on Code Arena, a real-world coding leaderboard, and posted strong scores on Agents' Last Exam, Vals Index, JobBench, and Chartography. It highlights coding, agentic tasks, and visual reasoning. On Work360, which Alibaba describes as a proprietary benchmark across 721 specialized fields, the company says Qwen4-Max outperforms Qwen3.8-Max overall. Those rankings are the vendor’s claims. They are not independently checked here.

The release also describes long agent runs. Alibaba says Qwen4 completed a continuous 48-hour run through four iterations of a closed loop—molecular design, synthesis and testing, and active learning—and discovered an experimentally validated coolant molecule aimed at next-generation chip cooling. In an Alzheimer’s research experiment, Alibaba says Qwen4-Max executed 454 tool calls and 3,239 metric evaluations to design and optimize 24 candidate molecules. In an automotive manufacturing showcase, the company says manual workpiece programming time fell by 95%, from roughly 475 minutes to 24 minutes, by catching overcuts and insufficient material removal before machining. Alibaba also says cloud resource scheduling tests cut data-annotation model call costs by an average of 9% and accelerated offline CPU rescheduling by more than 14 times.

TechNode notes a separate plugin that lets agents deliver an initial version of their work and revise it as users give feedback.

On safety, Alibaba says training included code security checks alongside functionality tests, that cybersecurity data was filtered to retain defensive material for vulnerability detection and remediation while removing offensive attack content, and that dedicated safety datasets and post-training target chemical, biological, radiological, and nuclear risks. That is the company’s account of its own governance, not a third-party audit.

What to read and try

Read Alibaba’s November 16 press release for the published size, the million-token window, and the cache-hit price, then TechNode for the November 11 preview-to-availability sequence and the Work360 figures in yuan. If you are calling models for coding or workplace agents, try Qwen4-Max on QwenCloud’s Token Plan or API and run your own evals before treating the Code Arena rank as settled. Do not plan on weights for this flagship; the open-weight promise in both accounts is for smaller Qwen4 models still to come.

Sources