Qwen3.5-9B

Qwen3.5-9B is the small-size workhorse of the Qwen3.5 generation, published by Alibaba in late February 2026 and by a wide margin the most downloaded model the Qwen team currently ships — roughly 12.6 million pulls from Hugging Face in a 30-day window, more than any single Qwen3.6 or Qwen3.8 checkpoint. The defining change in this generation is that vision is no longer bolted on. Earlier Qwen releases kept a text line and a separate -VL line; Qwen3.5 is trained from the start on mixed text and image tokens, and Alibaba's own numbers show the 9B model beating the previous dedicated vision model Qwen3-VL-30B-A3B across the board despite being a third of its size — MathVision 78.9 against 65.7, MMMU-Pro 70.1 against 63.0, and 93.7 against 72.5 on VlmsAreBlind, a test built from images that language priors cannot guess. On text it holds a similar position. Against OpenAI's GPT-OSS-120B, a model thirteen times larger, the vendor reports higher scores on MMLU-Pro (82.5 vs 80.8), GPQA Diamond (81.7 vs 80.1), IFEval (91.5 vs 88.9) and long-context retrieval (AA-LCR 63.0 vs 50.7). It gives ground where raw capacity still tells: LiveCodeBench v6 65.6 against 82.7, and competition mathematics (HMMT) by roughly seven points. All of these are vendor-reported figures. Architecturally it is a hybrid: 32 layers arranged as eight repetitions of three Gated DeltaNet blocks followed by one gated-attention block, hidden dimension 4,096, vocabulary 248,320 tokens, trained with multi-token prediction. Linear attention in three layers out of four is what keeps memory use bounded across a context window of 262,144 tokens natively, which Alibaba says extends to 1,010,000. Thinking mode is on by default and can be disabled per request. For readers outside the English-speaking world the relevant claim is coverage of 201 languages and dialects. Alibaba does not publish that list, so whether Polish is treated as a first-class language cannot be verified from the model card; what the card does report are aggregate multilingual scores — 81.2 on MMMLU, 76.3 on MMLU-ProX averaged over 29 languages, and 72.6 on a 55-language translation benchmark. The weights are Apache 2.0, with the full licence text present in the repository and no additional use restrictions attached.

▸Videos

No videos yet.

▸Reviews