Qwen3-VL Technical Report

TL;DR

Qwen3-VL is Alibaba Qwen’s 2025 vision-language model family, released as a coordinated multi-checkpoint package (2B / 4B / 8B / 32B dense + 30B-A3B / 235B-A22B MoE, each with Instruct and Thinking variants) and now documented in this November-27 technical report. The architecture introduces three changes over Qwen2.5-VL — Interleaved-MRoPE for video positional encoding, DeepStack multi-level ViT feature fusion, and explicit Text-Timestamp Alignment for video event localization — and the report covers pretraining, post-training, architecture, infrastructure, data, and evaluation in one document. At filing time the paper PDF was not retrievable through the fetcher (empty body); claims below are sourced from the companion GitHub README (Qwen3-VL — Multimodal LLM series (Dense + MoE, Instruct + Thinking)) and the public arXiv abstract page, which is why this page is unusually thin on quantitative numbers.

Key claims

Architecture vs Qwen2.5-VL introduces three changes: Interleaved-MRoPE (full-frequency allocation across time/width/height for video), DeepStack (multi-level ViT feature fusion into the LLM), and Text-Timestamp Alignment for video event localization that replaces T-RoPE [companion README §Model Architecture Updates, paper title/abstract].
The family ships as one report covering all model scales — 2B / 4B / 8B / 32B dense and 30B-A3B / 235B-A22B MoE, each with Instruct + Thinking variants — rather than per-checkpoint papers [paper abstract / tweet].
Native context length is 256K tokens, expandable to 1M, with second-level video event indexing claimed across hours-long video [companion README §Key Enhancements].
Visual capabilities include 32-language OCR (up from 10 in Qwen2.5-VL), 2D + 3D grounding, and an explicit “Visual Agent” capability with mobile and computer-use cookbooks shipped alongside the release [companion README §Key Enhancements, §Cookbooks].
The 8B model is the most-downloaded variant by a wide margin — 2M+ downloads in roughly one month according to the announcement [tweet copy].
The image processor uses a 32× spatial compression ratio (vs 14 for Qwen2.5-VL); the video processor adds 2× temporal compression on top [companion README §Pixel Control via Official Processor].
Headline benchmark numbers ship as PNG tables in the GitHub README and (presumably) full tables in the PDF — neither transcribed here because the PDF body was not retrievable at filing time [companion README §Performance images, paper PDF].

Method

The report covers pretraining, post-training, architecture, infrastructure, data, and evaluation; specifics beyond what the companion repo README already documents are not extractable from the empty-body PDF fetch at filing time. The three architecture levers are documented at the README level: Interleaved-MRoPE spreads rotary frequencies across the time / height / width axes for video tokens rather than partitioning them, which is the team’s bid for better long-horizon temporal information retention. DeepStack fuses features from multiple ViT layers into the LLM rather than only the final layer, preserving fine-grained visual detail that ViT pooling would otherwise lose. Text-Timestamp Alignment replaces T-RoPE-style positional encoding for video with an explicit alignment between text tokens and frame timestamps, enabling second-level event localization without relying on positional encoding alone to encode time. The release covers six parameter scales (2B / 4B / 8B / 32B dense; 30B-A3B / 235B-A22B MoE), each shipped as an {size}-Instruct and {size}-Thinking pair, with FP8 quantized variants for the MoE flagships. See Qwen3-VL — Multimodal LLM series (Dense + MoE, Instruct + Thinking) for the architecture and capability checklist; the tech report adds pretraining-data, post-training-recipe, and infra details that are not yet legible from the wiki’s snapshot of the PDF.

Results

The PDF body was not retrievable through the fetcher at filing time; the README ships performance tables as PNGs rather than transcribed numbers. Concrete benchmark figures are deferred to a /bud refresh once the report is parseable.
Family-level engagement signal: the announcement reports that 3 models passed 1M downloads in roughly one month, with Qwen3-VL-8B at 2M+ — concrete take-up data that the per-model wiki pages didn’t have at their filing time [tweet copy].

Why it’s interesting

This is the long-form companion to the umbrella release page Qwen3-VL — Multimodal LLM series (Dense + MoE, Instruct + Thinking) and the per-checkpoint announcements (Qwen3-VL compact dense — 4B and 8B (Instruct + Thinking), Qwen3-VL-2B and Qwen3-VL-32B dense — Instruct + Thinking, FP8 variants (Alibaba Qwen), Qwen3-VL-30B-A3B-Instruct, Qwen3-VL-30B-A3B-Thinking) — a single technical report covering an entire shipped model family is the same packaging convention the wiki already tracked for Qwen3-VL-Embedding and Qwen3-VL-Reranker: A Unified Framework for State-of-the-Art Multimodal Retrieval and Ranking (the Qwen3-VL-Embedding+Reranker joint report) and Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation (Qwen3-TTS), and reinforces a pattern documented under Open foundation-model releases: ship the full series + multi-backend serving + a single joint paper. Against the contemporaneous open-VLM cohort — Ovis2.5 Technical Report, Ovis-Image Technical Report, Kwai Keye-VL Technical Report, GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning, SAIL-VL2 Technical Report — Qwen3-VL is the broadest in checkpoint-scale coverage (six parameter scales, two variants each, plus quantization). The three architecture levers (Interleaved-MRoPE / DeepStack / Text-Timestamp Alignment) are the most concrete novelty and are also the bits most worth comparing against Scaling Spatial Intelligence with Multimodal Foundation Models (SAIL-VL2 / spatial-intelligence VLM scaling) and Cambrian-S’s spatial-supersensing position (Cambrian-S: Towards Spatial Supersensing in Video).