Alibaba launches Qwen4-Max with hybrid architecture and lower costs

Alibaba has released Qwen4-Max, a 2.4-trillion-parameter mixture-of-experts model designed for efficient long-context processing and autonomous agent tasks.

2 min read

Alibaba launches Qwen4-Max with hybrid architecture and lower costs

Alibaba has introduced Qwen4-Max, a new flagship artificial intelligence model designed to handle complex professional work in industry, science, and finance. The model features a hybrid architecture intended to process long texts more efficiently and at a lower cost than previous generations.

Key facts

Total parameters
2.4 trillion
Active parameters
71 billion
Context window
1 million tokens
Input price
$1.50 per 1M tokens
Output price
$4.50 per 1M tokens
Cache hit price
$0.075
Model Comparison on Intelligence Index and Task Cost
Intelligence Index Score
Qwen4-Max: 43
Qwen3.8-Max: 34
Kimi K3.1 Pro: 46
Tencent Hy4: 35
Cost per Task
Qwen4-Max: $3.27
Qwen3.8-Max: $10.02
Kimi K3.1 Pro: $3.68
Tencent Hy4: $1.03
Total Parameters
Qwen4-Max: 2.4 trillion
Qwen3.8-Max: 2.4 trillion
Kimi K3.1 Pro: 2.8 trillion
Tencent Hy4: —
Bar charts comparing Qwen4-Max performance across multiple benchmarks.
Benchmark comparisons across 16 tasks position Qwen4-Max against competing industry models., Source ↗

Architecture and efficiency

Qwen4-Max is a mixture-of-experts model — an architecture where the model routes each input token to specialized sub-networks rather than using the entire network at once. The model contains 2.4 trillion total parameters, with 71 billion active per token. It processes images natively and includes a one-million-token context window.

Diagram showing Qwen4-Max hybrid attention architecture and mixture of experts layers.
Architecture diagram detailing Qwen4-Max hybrid attention layers, Gated DeltaNet blocks, and MoE routing., Source ↗

To reduce the high computational cost of long inputs, the model uses a hybrid attention design. Three layers compress prior context into a condensed state, and a fourth layer selects only the relevant parts. This approach lowers compute costs because the model does not need to give every part of the context equal attention. An N-gram embedding layer with 328 billion parameters supplies common token sequences as ready-made building blocks.

Benchmarks and operating costs

On the Artificial Analysis Intelligence Index, Qwen4-Max ranks second among Chinese models behind Kimi K3.1 Pro, scoring 43 points at its highest reasoning setting. Operating costs are lower than its predecessor, with Artificial Analysis estimating the cost at $3.27 per index task, down from $10.02 for Qwen3.8-Max. Alibaba charges $1.50 per million input tokens and $4.50 per million output tokens through its API, with cache hits costing $0.075.

Agent capabilities and tools

The release includes a feature called Live Delivery, which allows users to inspect intermediate results from autonomous agents early and adjust running tasks on the fly. Alibaba has open-sourced the accompanying Qwen-Live-Delivery-Plugin. In practical demonstrations, the model has been paired with automated lab systems to synthesize chip coolants and design candidate drug molecules, and it can generate playable games and 3D models from natural-language descriptions.

Scene from a game generated by Qwen4-Max featuring a cat in a wizard outfit inside a medieval hall.
A playable 3D scene generated via natural language description using Qwen4-Max., Source ↗

Source: Original announcement

Worth a read?

Report

Comments