Alibaba releases natively multimodal Qwen3.5

Background

Alibaba's Qwen team had released its Qwen3 generation before this, and Qwen3.5 was the next generation of the family. Its design treats vision and language as one foundation rather than as a language model with a vision module added afterward. ^[1]

What happened

Alibaba released Qwen3.5 in February 2026. The flagship open model, Qwen3.5-397B-A17B, was posted to Hugging Face on 16 February 2026. Smaller open models followed between 24 and 28 February, namely 122B-A10B, 35B-A3B, 27B, 9B, 4B, 2B and 0.8B. ^[2][3]

Key details

The flagship has 397 billion total parameters and 17 billion activated per token. It uses 60 layers that combine Gated DeltaNet and Gated Attention blocks, with a 512-expert mixture-of-experts layer in which 10 routed experts and 1 shared expert are active. The weights are released under Apache 2.0. ^[1]

Alibaba reported 87.8 on MMLU-Pro and 91.3 on AIME26. The hosted Qwen3.5-Plus has a 1-million-token default context and built-in tools, and Alibaba said the model can operate desktop and mobile apps. ^[1][3]

Impact

The release is significant because it combined multimodal input, long context and agent-style tool use in a single open model family.

Several terms in the release are worth defining. Natively multimodal means that vision and language are learned together from the start, rather than attaching a vision encoder to a finished language model. A mixture-of-experts model activates only some of its parameters for each token, which is why the flagship's total size and active size differ. Agent-style tool use means the model can call external tools, such as search or code execution, while completing a task.

A mixture-of-experts model activates only part of its parameters for each token, which is why the total and active sizes in the flagship's name differ. Linear attention, the mechanism behind the Gated DeltaNet layers, is a cheaper approximation of full attention whose cost grows more gently with the length of the input.