Qwen4-Max is built for work that keeps going.
Alibaba’s flagship Qwen4 model combines native vision, long-context reasoning and tool use for coding, professional work and long-running agent tasks.
Released November 16, 2026- Context
- 1,000,000 tokens
- Input
- Text · image · video
- Reasoning
- Low · high · xhigh · off
A new flagship architecture.
The Qwen Team describes Qwen4-Max as a 2.4-trillion-parameter sparse mixture-of-experts model. It activates 71 billion backbone parameters per token and adds 328 billion parameters for N-gram embeddings.
The architecture combines Gated DeltaNet, Qwen Sparse Attention, gated residual streams and N-gram embeddings. The production API accepts text, images and video and returns text, with tool calling, parallel tool calls, structured output and prompt caching.
- Backbone parameters
- 2.4T
- Activated per token
- 71B
- Maximum output
- 131K standard · 262K reasoning
- Exact route
qwen/qwen4-max
What the published evaluations show.
These are Qwen Team results, not independent Minaco tests. The launch report combines Qwen’s internal evaluations with named public leaderboards.
Work360 is Qwen’s industry benchmark spanning 26 top-level fields, 182 subfields and 721 specialties. Read the official evaluation notes.
Minaco pricing.
Published Minaco rates in Credits per one million tokens.
- Input
- 3,000 Credits
- Cached input
- 150 Credits
- Cache write
- 3,750 Credits
- Output
- 9,000 Credits
Qwen4-Max is ready in Minaco.
Select it in Qwen Code or Hermes Agent, then work with the same local files, conversation history and permission controls used across Minaco Desktop.