On November 16, Alibaba's Qwen team officially released its flagship model Qwen4-Max, built on the all-new Qwen4 architecture: a 2.4T-parameter backbone, 71B activated per token, an additional 328B of N-gram Embedding parameters, and a 1M-token context window. The same day, the Agent Arena leaderboard put Qwen4-Max at #14 overall — the highest-ranked Chinese model — making Alibaba China's top lab and the world's #5 lab overall. This is the first model of the Qwen generation change, positioned by the company as a 24/7 AI teammate for long-horizon agent work, spanning 3D modeling, game development, finance, law, and healthcare.
What 2.4T parameters buy: fewer tokens, lower cost
The architecture upgrade centers on three changes: three Gated DeltaNet layers compress history into fixed-size states, with a lightweight indexer selecting context at micro-block granularity; four parallel residual branches use dynamic read and write gates to strengthen information flow between layers; and a second-layer N-gram Embedding expands capacity through context-based lookups. In-house evaluations show Qwen4-Max-Base leading the previous Qwen3.8-Max-Base on bits-per-byte across knowledge, code, QA, and STEM, down about 5% to 13% while activating 25% fewer parameters. On Work360, a benchmark for real professional work, it scored 52.69, up from 47.74 for the previous generation; average cost per task fell from $4.43 to $1.67, a 62% drop, while token usage on high-effort tasks fell about 24% to 35%, and about 48% to 49% on Terminal-Bench 4.0.
Long-horizon work is the selling point — but the benchmarks are the vendor's own
The most important thing for users is long-horizon execution. Qwen4-Max supports Live Delivery: tasks hand back a usable first version midway, then iterate on feedback, with users able to step in and correct course at any point, and the companion plugin is already open source. The showcased cases span a wide range: a single 48-hour continuous run produced a physical sample of a novel coolant material; financial research invoked 90 MCP tools and produced Excel workbooks and PowerPoint decks with traceable sources; 198 inconsistently formatted working papers were organized into an 86-page due-diligence report. The caveat is that these cases and benchmarks all come from the official release, with no independent third-party evaluation yet; research results like coolants and drug molecules are still far from everyday use, and most users should judge the model on office and coding work instead.
Where to use it now
Qwen4-Max is now live on QwenCloud, available through the Token Plan or the API, and also in Qwen Studio, QwenWork, and Qoder. The limited-time preview that opened on November 12 was only a prelude; the formal release fills in the architecture, benchmarks, and pricing. Two things are worth following next: when the smaller models in the same family go open source, as the Qwen team has previewed, and how independent evaluators check Work360 and other in-house benchmarks.