From SiliconANGLE · Maria Deutscher · · 1m
Alibaba Launches Qwen4-Max with Training Optimizations
Alibaba's Qwen4-Max model boosts output quality while cutting costs through training and attention upgrades.
In brief
Alibaba's Qwen4-Max model boosts output quality while cutting costs through training and attention upgrades. Alibaba released Qwen4-Max, an LLM with 2.4 trillion parameters that uses 35% fewer tokens than its predecessor for knowledge tasks. Originally reported by SiliconANGLE.
Alibaba Launches Qwen4-Max
Alibaba Group Holding Ltd. introduced Qwen4-Max, its newest and most capable large language model.
Higher Quality Output
Qwen4-Max generates higher-quality output than its predecessor while using up to 35% fewer tokens for knowledge tasks.
2.4 Trillion Parameters
Qwen4-Max’s backbone comprises 2.4 trillion parameters, activating only 71 billion at any given time.
N-gram Parameters
The model uses 328 billion N-gram parameters as a cache to save repeated calculations.
Residual Gradient Skipping
Qwen4-Max includes a residual feature that skips layers during training to reduce errors and costs.
Four-Channel Gradient Flow
The residual mechanism uses four gradient channels with a gate to adjust flow and improve training.
QSA Attention Optimization
Qwen4-Max’s attention mechanism skips unnecessary data to lower infrastructure costs for user prompts.
The main idea
Alibaba released Qwen4-Max, an LLM with 2.4 trillion parameters that uses 35% fewer tokens than its predecessor for knowledge tasks.
Read the TL;DR — this story on one card.

