Alibaba releases Qwen 3.5: sparse multimodal MoE, 262K context, Apache 2.0
In one sentence On 16 February 2026 Alibaba opened Qwen 3.5, starting with the 397B-A17B flagship: sparse MoE with 512 experts of which 11 are active, Gated DeltaNet attention, multimodal input, and 262,144 tokens of native context (around 1M with YaRN). Apache 2.0 and 201 languages; first among open-weight models on instruction following and multilinguality, still behind on reasoning.
On 16 February 2026 Alibaba published Qwen 3.5, the new generation of its open model family, under Apache 2.0: anyone can download, run, and modify these models, commercially included, at no cost.
The first model out was also the largest: 397 billion total parameters, but only 17 billion active on any given request. That is the Mixture of Experts principle — the model contains 512 specialists and only 11 are called for each question. The result is the quality of a very large model at the compute cost of a small one. The more manageable sizes followed within two weeks: 122B, 35B and 27B on 24 February, then 9B, 4B, 2B and 0.8B on 2 March — the ones you can genuinely run on a company server or a workstation.
Two characteristics matter more than the rest. First, it does not only read text: images and video go in too. Second, it can hold 262,144 tokens at once — several hundred pages in a single pass — and with an extension technique called YaRN that reaches roughly a million. It covers 201 languages.
On comparisons it pays to be precise. At launch the competitors were GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro. Qwen 3.5 comes out first among open models at following instructions and at multilingual work, but it stays behind on pure reasoning. So it is not the most capable model in absolute terms: it is the most useful one you can keep in-house.
Companies
Alibaba
Tools
Qwen 3.5
Tags
Sources