Qwen3.5-4B

Qwen3.5-4B is the smallest model of the Qwen3.5 generation that Alibaba still presents as a finished, deployable product rather than a research artefact. Published on 27 February 2026, it is pulled from Hugging Face roughly 7.3 million times in a 30-day window — second only to the 9B model within its own generation. The reason it matters to readers with ordinary hardware is arithmetic. The weight files hold 4.66 billion parameters in bfloat16, including the vision encoder, which is about 9.3 GB on disk — a single 12 GB or 16 GB consumer graphics card runs it, and quantised builds bring it below that. What that card then gives you is not a stripped-down text model but a full vision-language system: images, video frames and documents in, text out. Alibaba's own comparison is the striking part. Against Qwen3-VL-30B-A3B, the dedicated vision model of the previous generation, the 4B scores higher on MMMU (77.6 against 76.0), MMMU-Pro (66.3 against 63.0), MathVision (74.6 against 65.7) and VlmsAreBlind (92.6 against 72.5), a test built from images whose answers cannot be guessed from language alone. Against the hosted small models it is measured next to — GPT-5-Nano and Gemini 2.5 Flash-Lite — it leads on MMMU, MathVision, OCRBench and document understanding. On text it sits where its size suggests: MMLU-Pro 79.1, GPQA Diamond 76.2, IFEval 89.8, and a clear drop on code (LiveCodeBench v6 55.8) and competition mathematics against larger models. All figures are vendor-reported. Structurally it repeats the generation's hybrid layout: 32 layers arranged as eight repetitions of three Gated DeltaNet blocks and one gated-attention block, hidden dimension 2,560, vocabulary 248,320, output embeddings tied to the input embeddings, trained with multi-token prediction. Three linear-attention layers out of four are what keep memory bounded at a context window of 262,144 tokens, which the vendor says extends to 1,010,000 — a claim it makes for the 4B and larger, but not for the 2B and 0.8B below it. The licence is Apache 2.0, and the licence file in the repository is the unmodified Apache text with no added conditions.

▸Videos

No videos yet.

▸Reviews