Qwen3-VL — Multimodal LLM series (Dense + MoE, Instruct + Thinking)

TL;DR

Qwen3-VL is the umbrella release repo for Alibaba Qwen’s latest vision-language model family — initial drop (2025-09-23) shipped the Qwen3-VL-235B-A22B MoE in Instruct and Thinking variants, followed by 30B-A3B MoE (2025-10-04), 4B/8B dense (2025-10-15), and 2B/32B dense (2025-10-21). The series advertises native 256K context (expandable to 1M), 32-language OCR, 2D + 3D visual grounding, and a tech report (arXiv:2511.21631) released 2025-11-27. Architecture changes vs Qwen2.5-VL: Interleaved-MRoPE for video, DeepStack multi-level ViT feature fusion, and explicit Text-Timestamp Alignment for video temporal modeling.

Key claims

Initial release ships only the MoE flagship — Qwen3-VL-235B-A22B in Instruct and Thinking variants, with FP8 quantized counterparts on HF/ModelScope on day-0 [README §News 2025.09.23].
Architecture introduces three changes vs Qwen2.5-VL: Interleaved-MRoPE (full-frequency allocation across time/width/height), DeepStack (multi-level ViT feature fusion), and Text-Timestamp Alignment for video event localization (replaces T-RoPE) [README §Model Architecture Updates].
Native context length is 256K tokens, expandable to 1M, with second-level video indexing claimed across hours-long video [README §Key Enhancements].
32-language OCR (up from 10 in prior generation), with robustness claims under low light / blur / tilt and improved long-document structure parsing [README §Key Enhancements].
3D grounding is explicitly listed as a new capability alongside 2D grounding, intended for spatial reasoning and embodied AI [README §Key Enhancements, cookbooks/3d_grounding.ipynb].
Available in both Dense and MoE architectures spanning 2B / 4B / 8B / 32B dense and 30B-A3B / 235B-A22B MoE — same Instruct + Thinking-variant split across all sizes [README §News 2025.09.23 through 2025.10.21].
A separate Qwen3-VL technical report (arXiv:2511.21631) was released 2025-11-27, two months after the umbrella release began [README §News 2025.11.27].
Visual Agent capability is shipped with companion cookbooks for both mobile (mobile_agent.ipynb) and computer-use (computer_use.ipynb) settings, framing PC/mobile GUI control as a first-class capability [README §Cookbooks].

Method

Qwen3-VL keeps the Qwen2.5-VL split (ViT vision encoder → LLM backbone) but threads three changes through the multimodal interface. Interleaved-MRoPE allocates RoPE frequencies across the time / height / width axes for video tokens rather than partitioning them, intended to preserve long-horizon temporal information. DeepStack fuses features from multiple ViT layers (not just the last) into the LLM, aiming to retain fine-grained visual detail that gets washed out by ViT pooling. Text-Timestamp Alignment replaces the T-RoPE-style positional encoding for video with an explicit alignment between text tokens and frame timestamps, enabling second-level event localization rather than relying purely on positional encoding to encode time.

The release is structured as a model family rather than a single checkpoint: each parameter scale ships as {size}-Instruct and {size}-Thinking pair, and the MoE flagships also ship FP8-quantized variants. The repo provides per-capability cookbooks (omni recognition, document parsing, 2D and 3D grounding, OCR, video understanding, mobile/computer-use agents, thinking-with-images, multimodal coding, long-document understanding, spatial understanding) — the explicit framing is that each capability gets its own demonstration notebook rather than being implicit in benchmark numbers.

Results

Performance tables shipped in the README split the family along (Instruct vs Thinking) × (235B-A22B vs 30B-A3B vs 2B+32B dense vs 4B+8B dense) and report numbers on both visual and text-centric benchmarks [README §Performance images].
Image processor compression ratio is 32× spatial (vs 14 for Qwen2.5-VL); video processor compresses 32× spatial × 2× temporal, with per-modality min_pixels / max_pixels budgets configurable separately [README §Pixel Control via Official Processor].
Numerical headline scores are not legible from the README (delivered as PNG tables) so are not transcribed here; refer to the README images and the tech report (arXiv:2511.21631) for benchmark numbers.

Why it’s interesting

This is the umbrella release entry that ties together the Qwen3-VL artifacts already in the wiki: the day-0 dense-compact announcement (Qwen3-VL compact dense — 4B and 8B (Instruct + Thinking)) for 4B/8B, the 30B-A3B Instruct/Thinking blog cards (Qwen3-VL-30B-A3B-Instruct, Qwen3-VL-30B-A3B-Thinking), and the companion embedding/reranker series (Qwen3-VL-Embedding-8B, Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking) — confirming the pattern documented under Open foundation-model releases that a Qwen series ships as a coordinated multi-checkpoint package rather than per-model releases. It also sits next to the contemporaneous open-VLM cohort: Ovis2.5 Technical Report and the Ovis-Image follow-up (Ovis-Image Technical Report), Kwai’s Kwai Keye-VL Technical Report, the GLM-4.5V family (GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning), and SAIL-VL2 (SAIL-VL2 Technical Report) — all 2025 open VLM releases positioning against closed Gemini / GPT-4o-class models. The three architecture levers Qwen3-VL ships (Interleaved-MRoPE, DeepStack, Text-Timestamp Alignment) are the most concrete novelty here; the tech report at arXiv:2511.21631 is the right next read for benchmark-level comparison.