Alibaba Launches Qwen4-Max with 2.4T Parameters and Tokyo Region API Pricing
·2026.11.16 21:31
Key point
The new MoE model reduces active parameters to 71 billion while improving prediction accuracy, with Tokyo region input pricing at $1.413 per million tokens.
Details
- Alibaba released Qwen4-Max on November 16, featuring a 2.4-trillion-parameter MoE architecture that activates 71 billion parameters per token, a 25% reduction from the previous generation.
- The model improves internal BPB prediction accuracy by 5% to 13% while reducing cost per task on the Work360 benchmark by approximately 62%, from $4.43 to $1.67.
- API access is available in the Tokyo region (ap-northeast-1) with input pricing at $1.413 per million tokens and output pricing at $4.24 per million tokens.
- Independent evaluation by Artificial Analysis assigns an Intelligence Index score of 43, placing the model 38th among 81 evaluated models, below the top group scoring 55 to 65.
Architecture and Performance
- Uses GDN+QSA (Gated DeltaNet plus Qwen Sparse Attention) to compress past inputs into a fixed-size state, reducing computational cost for long inputs.
- Incorporates 328 billion parameters for N-gram Embedding to capture local text patterns and Gated Residuals to regulate information flow between layers.
- Supports up to 991,000 input tokens and 131,000 output tokens, or 983,000 input and 262,000 output tokens with thinking mode enabled.
Availability and Limitations
- Tokyo region supports real-time model calls only; batch processing, fine-tuning, and evaluation features are restricted to the Singapore region.
- Users can select a "Japan" deployment scope to restrict data processing to Japan, or "Global" to allow processing on any node including mainland China.
- Model weights for Qwen4-Max have not been released, though smaller Qwen4 models are planned for open-weight release.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.