Qwen3-VL-30B-A3B-Instruct

TL;DR

Open-weights Qwen3-VL MoE vision-language model with 30B total / 3B activated parameters, released in the Instruct (non-reasoning) post-training variant. Same vision-language stack as the larger Qwen3-VL-235B-A22B but at a much smaller activation cost — positioned as the “small” entry in the Qwen3-VL family, intended to run on a single consumer-class GPU while keeping the architectural upgrades (Interleaved-MRoPE, DeepStack multi-level ViT fusion, Text–Timestamp Alignment). Native 256K context expandable to 1M, 32-language OCR, 2D + 3D grounding. The sibling Qwen3-VL-30B-A3B-Thinking is the reasoning-tuned post-training of the same base.

Key claims

• 30B total parameters / 3B activated via MoE (the “A3B” suffix), released as Qwen/Qwen3-VL-30B-A3B-Instruct on Hugging Face with Qwen3VLMoeForConditionalGeneration as the model class [model card §Quickstart].
• Architecture introduces three named upgrades over Qwen2.5-VL: Interleaved-MRoPE (full-frequency positional embeddings allocated over time, width, and height for long-horizon video reasoning); DeepStack (fuses multi-level ViT features rather than only the final layer to capture fine-grained details); Text–Timestamp Alignment (replaces T-RoPE with timestamp-grounded event localization for video) [model card §Model Architecture Updates].
• Native 256K context, expandable to 1M tokens [model card §Key Enhancements].
• OCR coverage expanded from 19 → 32 languages, with claimed robustness to low light, blur, and tilt [model card §Key Enhancements].
• Adds 3D grounding on top of 2D grounding for spatial reasoning and embodied AI [model card §Key Enhancements].
• Visual-agent capabilities: claims operating PC/mobile GUIs (recognizing elements, invoking tools, completing tasks) and generating Draw.io/HTML/CSS/JS from images/videos [model card §Key Enhancements].
• Day-0 inference backends documented on the card: HF Transformers, vLLM, SGLang, Docker Model Runner [model card §Local Apps].
• ~1.2M downloads in the most recent month at filing time, suggesting heavy community adoption [model card §Downloads last month].

Method

The model card itself describes only the architectural deltas at a high level — for training details the card cites the Qwen3 Technical Report (arXiv:2505.09388), the Qwen2.5-VL Technical Report (arXiv:2502.13923), the Qwen2-VL paper (arXiv:2409.12191), and the original Qwen-VL paper (arXiv:2308.12966). Inference uses Qwen3VLMoeForConditionalGeneration from a recent build of HF Transformers (≥4.57.0); the card recommends attn_implementation="flash_attention_2" for multi-image and video. Inputs follow the standard Qwen-VL chat template with {"type": "image", "image": ...} and {"type": "text", "text": ...} content blocks. No new benchmark numbers are reported on the card itself beyond the “Multimodal performance” header — the model is positioned by association with the broader Qwen3-VL release rather than a per-checkpoint scorecard.

Results

The card does not enumerate per-benchmark numbers for the 30B-A3B variant specifically — that section is left as a placeholder (“Multimodal performance” header with no table in the fetched content). The headline framing is qualitative: “the most powerful vision-language model in the Qwen series to date” and the Dense + MoE × Instruct + Thinking matrix covering “edge to cloud.” Downstream evidence of capability comes via the derivative releases — Qwen3-VL-Embedding-8B and Qwen3-VL-Reranker-8B (Qwen3-VL-Embedding-8B, Qwen3-VL-Reranker-8B) are fine-tunes of Qwen3-VL-8B-Instruct that reach 77.9 MMEB-V2 overall, a useful indirect floor on the base VLM’s representation quality.

Why it’s interesting

This is the smallest active-parameter Qwen3-VL release (3B active), which makes it the natural target for downstream specialization — the same role Qwen3-VL-8B-Instruct played as the base for Qwen3-VL-Embedding-8B and the role Qwen2.5-VL-8B played as the base for Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning. Pairing this Instruct variant with its Qwen3-VL-30B-A3B-Thinking sibling continues the now-standard practice of shipping reasoning- and non-reasoning-tuned post-training variants together (the same pattern observed in MiniMax-M2.5 and GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models). The 3B-active MoE design slots it into MoE Routing Design as the smallest active-count VLM tracked here — useful for downstream teams considering whether the Qwen3-VL stack can serve as an internal multimodal backbone without a frontier-scale serving budget.

See also

• Qwen3-VL-30B-A3B-Thinking — sibling reasoning-tuned variant of the same 30B-A3B base.
• Qwen3-VL-Embedding-8B — Qwen3-VL-8B-Instruct retrained as a multimodal embedding model; same family, smaller sibling.
• Qwen3-VL-Reranker-8B — reranker counterpart of the embedding model.
• Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking — joint technical report covering the Qwen3-VL embedding/reranker series; references the base Qwen3-VL VLMs.
• Open foundation-model releases — broader pattern of coordinated open-weights releases this fits into.
• MoE Routing Design — the A3B (3B-active) MoE design choice.
• Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning — out-of-house attempt to add reasoning to an older Qwen-VL base via RL; the Thinking sibling is the in-house analogue on a newer base.