Qwen3.5: open weights, 397B
On 15-16 February 2026 Qwen (Alibaba) released Qwen3.5-397B-A17B, the first model of the series with open weights: native vision and language, 17 billion active of 397 billion parameters, 201 languages, Apache 2.0 licence.
Why it matters
A large multimodal model with a hybrid architecture became available with weights under a permissive licence. An editorial assessment: the comparison table is the developer's.
What is said. Per the post, native vision-language, a hybrid of linear attention (Gated Delta Networks) with a sparse mixture of experts; 201 languages and dialects instead of 119. Model card: 60 layers, 512 experts (10 routed and 1 shared), 262,144-token context extensible to 1,010,000. The hosted Qwen3.5-Plus has a 1-million-token context by default and built-in tools. Apache 2.0 licence. What the record does not claim. The comparison table against GPT-5.2, Claude 4.5 Opus, Gemini-3 Pro and others is Qwen's own measurement; none of it was checked. The 1 million context belongs to the hosted version, not the weights.