Qwen3-VL-235B-A22B-Thinking: The Reasoning Edition of Qwen3-VL

What this model is

Qwen3-VL-235B-A22B-Thinking is the reasoning-enhanced edition of Alibaba's largest vision-language model: 235 billion total parameters with 22 billion active per token, Apache 2.0 and self-hostable, with a native 256K token context expandable to 1M. Alibaba publish the Qwen3-VL line in dense and mixture-of-experts architectures scaling from edge to cloud, each in an Instruct and a Thinking edition. Alibaba position the Thinking edition for multi-step visual reasoning.

LMArena

On LMArena's text leaderboard Qwen3 VL 235B A22b Thinking has 1395 Elo (rank 171, thinking effort); by category, 1452 Elo in Coding (rank 158, thinking effort), 1399 Elo in Math (rank 166, thinking effort), 1441 Elo in Expert (rank 136, thinking effort) and 1338 Elo in Creative Writing (rank 195, thinking effort). The Instruct edition has 1413 Elo (rank 148). LMArena's text board does not show the model an image, so it does not measure visual reasoning.

Price and successor

Roughly $0.40 per million input tokens and $4 output through third-party providers; reasoning tokens bill as output. Alibaba state that the Qwen3.5 generation outperforms the Qwen3-VL line across visual understanding.

What Alibaba list for Qwen3-VL

Alibaba list the Qwen3-VL generation's capabilities as: a visual agent that operates PC and mobile interfaces, recognising elements, understanding functions, invoking tools and completing tasks; visual coding that generates Draw.io diagrams and HTML, CSS and JavaScript from images and video; spatial perception covering object positions, viewpoints and occlusion, with 2D grounding and 3D grounding; long-context and video understanding, with a native 256K context expandable to 1M and second-level indexing over hours-long video; broader visual recognition of celebrities, anime, products, landmarks, plants and animals; and OCR in 32 languages, up from 19, with better handling of low light, blur, tilt, rare characters and long-document structure. For the Thinking edition Alibaba highlight multimodal reasoning in STEM and mathematics, including causal analysis and evidence-based answers.

Architecture changes

Alibaba describe three architectural changes. Interleaved-MRoPE allocates positional frequencies across time, width and height for long-horizon video reasoning. DeepStack fuses multi-level ViT features to capture fine detail and sharpen image-text alignment. Text-timestamp alignment replaces T-RoPE with timestamp-grounded event localisation for video. The model loads through the Qwen3VLMoeForConditionalGeneration class in Transformers, which Alibaba advised building from source before version 4.57.0 was released, and Alibaba recommend flash_attention_2 for multi-image and video inputs.

Release and settings

Alibaba released Qwen3-VL-235B-A22B in Instruct and Thinking editions on 23 September 2025. The 30B-A3B models and FP8 versions across the collection followed on 4 October, the 4B and 8B models on 15 October, and the 2B and 32B models on 21 October, each in Instruct and Thinking editions; the technical paper was published on 27 November 2025. Qwen3-VL is supported in vLLM 0.11.0 and later, and Alibaba's efficient H100 and H200 deployment uses tensor parallelism of 8 with expert parallelism enabled. For Thinking models Alibaba recommend temperature 0.6, top_p 0.95, top_k 20, a presence penalty of 0 and up to 40,960 output tokens.

Lineage

Qwen3-VL follows Qwen2.5-VL, Qwen2-VL and the original Qwen-VL, each with a technical report on arXiv. The model can be loaded through ModelScope as well as Hugging Face, and Alibaba publish separate vision and text benchmark tables for the Thinking editions.

Official Resources