Skip to content
Neural Daily

ModelsNews

Qwen4-Max: 2.4T MoE, 1M context, $1.50 input

🇭🇺 Olvasd el magyarul →

Alibaba has released Qwen4-Max, the first model of its Qwen4 generation. The 2.4-trillion-parameter MoE activates 71 billion parameters per token, has native vision and a 1M-token context window. It is API-only for now; smaller open-weight Qwen4 models are promised later.

· 2 min read

A Qwen bejelentő grafikája: Qwen 4 Is Now Available felirat, mellette egy pagodaszerű épület vázlatos rajza
Image: Alibaba Cloud · Press kit · Source

What shipped?

Qwen4-Max is Alibaba's most capable model so far (Alibaba Cloud). It is a sparse MoE with 2.4 trillion total parameters and 71 billion active per token, 25% fewer active parameters than Qwen3.8-Max at the same size. The new architecture reworks attention, residual connections and embeddings. It accepts text, images and video.

It is a closed model, available through the QwenCloud API and Token Plan, as well as in QwenWork and QwenStudio. Alibaba says smaller open-weight models from the Qwen4 series are coming soon, without giving a date.

What does it cost?

  • Input: $1.50, output: $4.50 per million tokens (
  • [object Object]
  • ).
  • Implicit and explicit cache reads: $0.075 per million tokens.
  • Context: 1M tokens, up to 991K input and 131K output tokens (262K in thinking mode).
  • Built-in tools: code interpreter, web search, Responses API support.

The input price matches Gemini 4 Hydrogen, while output is cheaper.

What do the numbers show?

Alibaba says it ranks sixth on Code Arena. On its in-house Work360 test, covering tasks from 721 fields, it beats Qwen3.8-Max while using 35% fewer tokens in high reasoning mode and cutting average cost per task by about 60%. These are vendor figures.

  • Agents' Last Exam v1.1: 46.7, ahead of Claude Opus 5.5 at 43.4 and GPT-6.1 Sol at 42.6 (Alibaba chart, via
  • [object Object]
  • ).
  • Terminal-Bench 4.0: 48.5. Vals confirmed this independently, but Kimi K3.1 (67.2) and Opus 5.5 (59.6) are well ahead here.
  • ProgramBench: 31.5 against 65.0 for Opus 5.5.
  • Independent indexes are more mixed: Artificial Analysis ranks it ninth, Vals eleventh.

Who should try it?

Qwen4-Max is strongest at long-running agentic work and document-style deliverables. Alibaba showed a 48-hour, four-iteration agent run that found an experimentally validated coolant molecule. For terminal coding and visual tasks the numbers say it is not the first pick yet. Self-hosters will have to wait for the promised smaller open-weight models.

Key takeaways

  • 2.4T-parameter MoE, 71B active, 1M context, native vision.
  • $1.50/$4.50 per million tokens, $0.075 cache reads.
  • Strong on agentic benchmarks, behind on terminal coding.
  • Closed for now; smaller open-weight Qwen4 models later.

Related articles