Alibaba launches Qwen3.5 which, with the open model Qwen3.5-397B-A17B, is committed to native multimodal agents

Alibaba has announced the launch of the Qwen3.5 serieswhich has the open artificial intelligence (AI) model Qwen3.5-397B-A17Bwith which it allows developers and companies to increase their productivity thanks to its 397,000 million parameters, betting on native multimodal agents.

Compared with its predecessor models, such as the Qwen 3 series introduced in April last year, the new Qwen3.5-397B-A17B model has achieved “outstanding results” in performance assessments, such as reasoning, coding, agentic abilities, and multimodal understanding.

This is because the new model is based on a hybrid architecture that combines linear information (through Gated Delta Networks) with various experts, reaching a total of 397 billion parametersof which 17,000 million are activated, thus optimizing its speed and cost, as the company explained in its blog. Also Alibaba has expanded the languages ​​and dialects it understands, going from 119 to 201.

To improve performance, when training the Qwen3.5 series models, Alibaba has focused on the Expanding RL Machine Learning Tasks and Environmentsincreasing the difficulty and generalizability of these environments rather than optimizing them for specific metrics or narrow topic queries.

That is, for example, it has been trained on a significantly larger scale of visual text tokens, compared to Qwen3, with rich multilingual, STEM and reasoning data, “under stricter filtering.” As a result, the company has claimed that the decoding performance of Qwen3.5-397B-A17B is 3.5 to 7.2 times higher than that of Qwen3-235B-A22B.

🚀 Qwen3.5-397B-A17B is here: The first open-weight model in the Qwen3.5 series.

🖼️Native multimodal. Trained for real-world agents.
✨Powered by hybrid linear attention + sparse MoE and large-scale RL environment scaling.
⚡8.6x–19.0x decoding throughput vs Qwen3-Max
🌍201… pic.twitter.com/Pq0qIk54MB

— Qwen (@Alibaba_Qwen) February 16, 2026

Thanks to your hybrid architecture and his native multimodal reasoningAlibaba has clarified that it seeks to compete with other companies in the sector such as OpenAI or Anthropic. Regarding the company’s future goals, they are focusing on creating “coherent systems that work autonomously,” evolving current assistants into constant and reliable partners, capable of executing complex objects with a “human-aligned criterion.”

With all this, Qwen3.5 is available in Qwen Chat in three modes: automatic, thinking and fast. With automatic mode, users will be able to take advantage of features such as adaptive thinking, search, and the code interpreter, while with thinking mode, the model analyzes complex problems in depth. For its part, quick mode answers questions instantly “without spending tokens on thinking.”