Qwen3-VL 8B

Qwen3-VL 8B is the dense 8B-parameter instruct variant of Alibaba's Qwen3-VL vision-language family (2B/4B/8B/32B dense + 30B-A3B/235B-A22B MoE), released under Apache 2.0. It takes interleaved text, images, and video as input, with a native 256K-token interleaved context window (provider-claimed expandable to 1M). Key upgrades over Qwen2.5-VL: interleaved-MRoPE positional embeddings, DeepStack multi-level ViT feature fusion, and text-based timestamp alignment for video.

Benchmarks (independent): a third-party evaluation of the base model (hongruijia DPE v3 table, Jan 2026) reports MMMU 65.44, MMVet 67.29, MMStar 61.27, MathVerse 53.22, MathVision 51.97, CharXiv 47.20, BLINK 68.54. An independent real-world test on 137 messy documents (vibingtalk, Sep 2026) scored Qwen3-VL 8B at 59% fully correct vs GPT-5.6-terra 57% — it beat the closed model on IRS tax forms (21/32 vs 7/32) but failed on Indian date formats (2/10). Provider-reported (not independently verified): Qwen claims the 8B model outperforms the much larger Qwen2.5-VL-32B on several multimodal benchmarks, and the full family leads on MMMU-class and visual-math evaluations.

With 8.6M+ Hugging Face downloads, it is one of the most-adopted open vision-language models in the world. Served via Alibaba Cloud Model Studio (DashScope, OpenAI-compatible) and self-hostable via vLLM/Transformers/Ollama.

Independent: MMMU 65.44, MMVet 67.29, MMStar 61.27, MathVerse 53.22, MathVision 51.97, CharXiv 47.20, BLINK 68.54 (third-party DPE v3 eval of base model, Jan 2026). Real-world messy-document test: 59% fully correct on 137 documents, beating GPT-5.6-terra (57%) — strong on tax forms, weak on Indian date formats.

Provider-reported: Outperforms Qwen2.5-VL-32B on several multimodal benchmarks (Alibaba, not independently verified).