We introduce Qwen3-VL, the most capable vision-language model in the Qwen series to date, achieving superior performance across a broad range of multimodal benchmarks. It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B) and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency-quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger pure-text understanding, surpassing comparable text-only backbones in several cases; (ii) robust long-context comprehension with a native 256K-token window for both text and interleaved multimodal inputs, enabling faithful retention, retrieval, and cross-referencing across long documents and videos; and (iii) advanced multimodal reasoning across single-image, multi-image, and video tasks, demonstrating leading performance on comprehensive evaluations such as MMMU and visual-math benchmarks (e.g., MathVista and MathVision). Architecturally, we introduce three key upgrades: (i) an enhanced interleaved-MRoPE for stronger spatial-temporal modeling across images and video; (ii) DeepStack integration, which effectively leverages multi-level ViT features to tighten vision-language alignment; and (iii) text-based time alignment for video, evolving from T-RoPE to explicit textual timestamp alignment for more precise temporal grounding. Under comparable token budgets and latency constraints, Qwen3-VL achieves superior performance in both dense and Mixture-of-Experts (MoE) architectures. We envision Qwen3-VL serving as a foundational engine for image-grounded reasoning, agentic decision-making, and multimodal code intelligence in real-world workflows.
核心贡献 · Key contributions
提出 Qwen3-VL 系列，包括密集（2B-32B）和 MoE（30B-A3B, 235B-A22B）变体，原生支持 256K 词元的交错多模态上下文。 Introduces Qwen3-VL family with dense (2B-32B) and MoE (30B-A3B, 235B-A22B) variants, natively supporting 256K-token interleaved multimodal contexts.
架构升级包括交错 MRoPE、用于跨层视觉-语言融合的 DeepStack，以及用于视频对齐的显式文本时间戳。 Architecture upgrades include interleaved MRoPE, DeepStack for cross-layer vision-language fusion, and explicit text-based timestamps for video alignment.
训练创新：平方根重加权以平衡文本和多模态损失，四阶段预训练达到 256K 上下文，后训练分为非思考和思考变体。 Training innovations: square-root reweighting to balance text and multimodal loss, four-stage pre-training up to 256K context, and bifurcated post-training into non-thinking and thinking variants.
大规模精选数据涵盖描述、OCR、文档、视频、空间推理、代码和 GUI 智能体，并采用严格的过滤和合成流程。 Massive curated data spanning captions, OCR, documents, video, spatial reasoning, code, and GUI agents, with rigorous filtering and synthesis pipelines.
在 MMMU、MathVista、MathVision 等多项基准上取得领先结果，覆盖非思考和思考模式。 State-of-the-art results on multiple benchmarks including MMMU, MathVista, MathVision, and visual reasoning tasks in both non-thinking and thinking modes.
在多模态提升的同时保持了强大的纯文本性能，在多个案例中超越同等的文本骨干模型。 Strong text-only performance preserved alongside multimodal gains, surpassing comparable text backbones in several cases.
局限 · Limitations
旗舰 235B-A22B 模型需要巨大算力，限制了在资源丰富环境外的部署和实时应用。 The flagship 235B-A22B model requires massive compute, limiting deployment to resource-rich environments and real-time applications.
依赖大规模合成数据和伪标签可能引入噪声，降低对未见过的真实世界分布的泛化能力。 Dependence on large-scale synthetic data and pseudo-labeling may introduce noise and reduce generalization to unseen real-world distributions.
在极长视频或文档上，256K 词元的长上下文理解仍可能因稀疏时间采样而性能下降。 Long-context understanding up to 256K tokens may still suffer from degraded performance on extremely long videos or documents due to sparse temporal sampling.
对齐中的详细失败模式和偏见（如语言混合、幻觉）未被完全表征；缓解依赖特定强化学习数据。 Detailed failure modes and biases in alignment (e.g., language mixing, hallucinations) are not fully characterized; mitigation relies on targeted RL data.
评估主要在静态基准上进行；实际智能体式和交互式性能可能不同，开放任务中的安全性未广泛验证。 Evaluation mainly on static benchmarks; real-world agentic and interactive performance may differ, and safety in open-ended tasks is not extensively validated.