Qwen 4 Max: a 2.4T-parameter teammate for long-running work

Qwen 4 Max is Alibaba's new flagship for professional work, creative production, and research. Announced by the Qwen team on November 16, 2026, Qwen4-Max combines a 1M context window, native vision and language, and a workflow for reviewing an AI agent's output while it is still working.

Qwen 4 Max: 1M context and a sparse architecture
Qwen4-Max uses a new sparse mixture-of-experts architecture with 2.4 trillion total parameters, activating 71 billion per token, plus 328 billion N-gram Embedding parameters. It combines Gated DeltaNet layers with micro-block sparse attention and four parallel Gated Residual branches.
The headline spec for large projects is the 1,000,000-token context window — room for an entire codebase or a full document archive in a single run.
Qwen reports that the base model reaches roughly 5–13% lower bits-per-byte than Qwen3.8-Max-Base across Knowledge, Code, QA, and STEM evaluations while activating 25% fewer parameters.
Benchmark results: better scores, fewer tokens
On Work360, a benchmark spanning 26 Tier-1 domains, 182 subdomains, and 721 specialized fields, Qwen4-Max scored 52.69, compared with 47.74 for Qwen3.8-Max, according to Qwen.
The team also reports lower token use: approximately 24% fewer tokens at Xhigh effort and 35% fewer at High on Work360, plus 48–49% fewer on Terminal-Bench 4.0.

Qwen's reported average Work360 cost per task fell from $4.43 to $1.67, a 62% reduction. The announcement says API pricing is below Qwen3.8-Max, particularly for cached input, but does not publish exact rates. These are vendor-reported results, not a guarantee for every workload.
An AI agent you can redirect mid-task
The most practical workflow change is Live Delivery: long-running agents stream intermediate deliverables so users can review progress and redirect the task before completion. Qwen has released an open-source Live Delivery plugin.
The demonstrations range from playable games built in Unity 6, Unreal Engine, and Godot to Blender modeling and agentic video production. Below is "Meow of Magic," a cat-mage adventure the model built in Unreal Engine.
Professional examples include financial research using 90 MCP tools to produce Excel workbooks and PowerPoint decks, and legal due diligence turning 198 working papers into a cited 86-page report. The pattern is the same everywhere: sustained tool use, then a deliverable you can review.
Where to try Qwen4-Max
Qwen lists availability through QwenCloud's Token Plan or API, plus QwenWork, Qoder, and QwenStudio. Qwen4-Max weights are not open; smaller Qwen4 models are promised for open-source release later.
You can try frontier models in vega.chat or through api.vega.chat; watch the model picker for Qwen 4 Max availability.
For medical, legal, financial, and other high-impact tasks, Qwen's own caveat applies: the model assists professional judgment. It does not replace it.