Alibaba's new AI model activates 3B parameters to challenge Gemma

Alibaba Cloud is intensifying the AI price war with a new model that activates just 3 billion parameters, a move targeting the high inference costs of large-scale systems from competitors like Google and Meta.

"The Qwen3.6-35B-A3B model utilizes a sparse Mixture-of-Experts, or MoE, architecture to achieve high efficiency," Alibaba's Tongyi Lab said in the April 16 announcement. The model has 35 billion total parameters but only activates a fraction for any given task.

This efficiency allows the new model to compete with larger dense models like Google’s recently released Gemma-31B while using significantly fewer computational resources. For developers, this translates directly to lower operational costs when running AI applications, a critical factor for widespread adoption. The model also shows significant improvements in programming tasks over its predecessor, the Qwen3.5-35B-A3B.

The release signals a direct challenge to other major AI players and could pressure margins across the sector. By open-sourcing the model, Alibaba (BABA) aims to accelerate its adoption within the developer community, potentially taking market share from incumbents like Nvidia (NVDA) on the hardware side and other model providers. The strategy hinges on creating a large ecosystem around its Qwen models, which are compatible with popular coding assistants including Claude Code and Qwen Code. This could bolster Alibaba Cloud's competitive position as it vies for a larger piece of the estimated $1 trillion AI market.

This article is for informational purposes only and does not constitute investment advice.