Alibaba launches Qwen 3.5 with 397 billion parameters

Alibaba has announced the launch of the AI model Qwen 3.5, featuring 397 billion parameters but activating only 17 billion per token, making it more economical and efficient. This model surpasses the previous Qwen3-Max, which had over one trillion parameters, and was designed to coincide with the Lunar New Year.

The Qwen 3.5 is an evolution of the experimental Qwen3-Next model, showcasing an architecture with 512 specialists that drastically improves inference latency. With a context length of 256k, Qwen 3.5 decodes up to 19 times faster than Qwen3-Max and 7.2 times faster than Qwen3 235B-A22B.

Economically, the new model is 60% cheaper to operate and can handle eight times larger concurrent workloads. Compared to Google’s Gemini 3 Pro model, Qwen 3.5 costs about 1/18 of the price. It also incorporates expanded multilingual coverage for 201 languages, further reducing inference costs.

Qwen 3.5 is natively multimodal, trained on text, images, and video, outperforming models that add visual components later. For businesses, the open version of Qwen 3.5 is available under the Apache 2.0 license, allowing unrestricted commercial use.

In conclusion, Qwen 3.5 represents a viable option for companies seeking cutting-edge AI models, with multimodal capabilities and a context window of 1 million tokens, without the need for proprietary APIs.