Alibaba has unveiled Qwen 3.5, a breakthrough in enterprise AI procurement that challenges the conventional model of renting AI infrastructure. The new flagship model, Qwen3.5-397B-A17B, boasts 397 billion total parameters but activates only 17 billion per token, outperforming Alibaba’s previous trillion-parameter model at a fraction of the cost.
Qwen3.5’s architecture, a successor to Qwen3-Next, features innovative engineering with 512 experts, leading to significantly lower inference latency and faster decoding speeds compared to previous models. Alibaba claims a 60% reduction in operational costs and increased workload handling capacity, positioning Qwen 3.5 as a cost-effective and high-performance option for AI deployments.
Moreover, Qwen3.5 integrates native multimodal capabilities, training on text, images, and video simultaneously for enhanced performance on tasks requiring tight text-image reasoning. With expanded multilingual support and improved tokenizer efficiency, the model offers global deployment advantages, reducing inference costs and improving response times for multilingual user bases.
Qwen3.5 also introduces agentic capabilities, enabling autonomous actions and complex coding tasks through the open-source Qwen Code interface. The model’s adaptive inference modes cater to diverse enterprise needs, balancing real-time interactions and deep analytical workflows efficiently.
With the release of Qwen 3.5, Alibaba sets the stage for a new era of enterprise AI procurement, offering open-weight models under the Apache 2.0 license for extensive commercial use. The industry anticipates further releases in the Qwen3.5 family, promising smaller dense distilled models and additional configurations to meet evolving AI demands.
Source: VentureBeat