Changelog

Release history and updates for Qwen3-VL

2025

November 27, 2025 - Technical Paper Release

Qwen3-VL Technical Paper Published

We released the Qwen3-VL technical paper, providing comprehensive technical details about the model architecture, training methodology, and evaluation results.

Key Topics Covered:

• Interleaved-MRoPE architecture
• DeepStack multi-level feature fusion
• Text-Timestamp alignment for videos
• Training data and methodology
• Comprehensive benchmark evaluations
• Ablation studies

Resources:

• Paper: arXiv:2511.21631
• Blog: Qwen AI Blog

---

October 21, 2025 - Qwen3-VL 2B & 32B Models

New Model Releases

Released four new models expanding the Qwen3-VL family:

• Qwen3-VL-2B-Instruct (HuggingFace)
• Qwen3-VL-2B-Thinking (HuggingFace)
• Qwen3-VL-32B-Instruct (HuggingFace)
• Qwen3-VL-32B-Thinking (HuggingFace)

Highlights:

• 2B models suitable for edge deployment and consumer GPUs
• 32B models provide high performance with reasonable resource requirements
• Both Instruct and Thinking editions available
• Available on both HuggingFace and ModelScope

---

October 15, 2025 - Qwen3-VL 4B & 8B Models

New Model Releases

Released four new models in the mid-size range:

• Qwen3-VL-4B-Instruct (HuggingFace)
• Qwen3-VL-4B-Thinking (HuggingFace)
• Qwen3-VL-8B-Instruct (HuggingFace)
• Qwen3-VL-8B-Thinking (HuggingFace)

Highlights:

• Balanced performance-to-resource ratio
• 4B suitable for RTX 3090/4090 GPUs
• 8B provides strong capabilities for professional use
• Both editions support full feature set

---

October 4, 2025 - MoE Models & FP8 Quantization

Qwen3-VL-30B-A3B Release

Released our first Mixture-of-Experts (MoE) models:

• Qwen3-VL-30B-A3B-Instruct (HuggingFace)
• Qwen3-VL-30B-A3B-Thinking (HuggingFace)

FP8 Quantized Models

Released FP8 quantized versions of all Qwen3-VL models for efficient deployment on H100/H200 GPUs.

Features:

• 30B total parameters, 3B active per token
• Efficient inference with MoE architecture
• FP8 models available in collections:
  • HuggingFace Collection
  • ModelScope Collection

---

September 23, 2025 - Qwen3-VL 235B Launch

Qwen3-VL-235B-A22B Release

Released the flagship Qwen3-VL models:

• Qwen3-VL-235B-A22B-Instruct (HuggingFace)
• Qwen3-VL-235B-A22B-Thinking (HuggingFace)

Major Features:

• State-of-the-art vision-language performance
• 235B parameters, 22B active (MoE)
• Native 256K context, expandable to 1M
• Advanced capabilities:
  • Visual agent (PC/mobile GUI interaction)
  • Visual coding (Draw.io, HTML/CSS/JS)
  • 3D grounding and spatial reasoning
  • Enhanced video understanding
  • 32-language OCR
  • Superior multimodal reasoning

Blog: Official announcement

---

April 8, 2025 - Fine-tuning Code Release

Fine-tuning Support

Released fine-tuning code for Qwen2-VL and Qwen2.5-VL, compatible with Qwen3-VL.

Resources:

• Code: qwen-vl-finetune
• Supports custom visual tasks
• LoRA and full fine-tuning

---

March 25, 2025 - Qwen2.5-VL-32B

Qwen2.5-VL-32B-Instruct Release

Released Qwen2.5-VL-32B-Instruct.

Improvements:

• Smarter responses
• Better human preference alignment
• Enhanced reasoning capabilities

Blog: Announcement

---

February 20, 2025 - Qwen2.5-VL Technical Report

Technical Report & AWQ Models

Released the Qwen2.5-VL Technical Report along with AWQ-quantized models:

• Qwen2.5-VL-3B-Instruct-AWQ
• Qwen2.5-VL-7B-Instruct-AWQ
• Qwen2.5-VL-72B-Instruct-AWQ

---

January 28, 2025 - Qwen2.5-VL Series

Qwen2.5-VL Family Release

Released the complete Qwen2.5-VL series on HuggingFace.

Models:

• Qwen2.5-VL-3B-Instruct
• Qwen2.5-VL-7B-Instruct
• Qwen2.5-VL-72B-Instruct

Blog: Announcement

---

2024

December 25, 2024 - QvQ-72B-Preview

Visual Reasoning Research Model

Released QvQ-72B-Preview, an experimental model focusing on enhanced visual reasoning.

Features:

• 72B parameters
• Advanced visual reasoning capabilities
• Research preview for community feedback

Blog: QvQ announcement

---

September 19, 2024 - Qwen2-VL-72B

Large Model & Quantization

Released Qwen2-VL-72B-Instruct with multiple quantized versions:

• Qwen2-VL-72B-Instruct
• Qwen2-VL-72B-Instruct-AWQ
• Qwen2-VL-72B-Instruct-GPTQ-Int4
• Qwen2-VL-72B-Instruct-GPTQ-Int8

Qwen2-VL Paper

Published Qwen2-VL technical paper.

---

August 30, 2024 - Qwen2-VL Series Launch

Qwen2-VL Family Release

Released the Qwen2-VL series:

• Qwen2-VL-2B-Instruct
• Qwen2-VL-7B-Instruct
• (72B announced, released later)

Features:

• Multi-resolution image understanding
• Video support
• Enhanced OCR
• Improved multimodal reasoning

Collection: Qwen2-VL

Blog: Qwen2-VL announcement

---

Release Timeline Summary

Date	Release	Highlights
2025-11-27	Technical Paper	Architectural details and evaluations
2025-10-21	2B & 32B Models	Edge to high-performance range
2025-10-15	4B & 8B Models	Mid-range balanced performance
2025-10-04	30B-A3B MoE & FP8	Efficient MoE, quantized models
2025-09-23	235B-A22B Flagship	State-of-the-art VLM
2025-04-08	Fine-tuning Code	Custom training support
2025-03-25	Qwen2.5-VL-32B	Improved alignment
2025-02-20	Qwen2.5-VL Report & AWQ	Technical documentation
2025-01-28	Qwen2.5-VL Series	Complete model family
2024-12-25	QvQ-72B Preview	Visual reasoning research
2024-09-19	Qwen2-VL-72B & Paper	Large model & documentation
2024-08-30	Qwen2-VL Launch	Series introduction

Migration Guides

From Qwen2.5-VL to Qwen3-VL

Key Changes:

1. Patch size: 14 → 16

   # Update qwen-vl-utils calls
   process_vision_info(messages, image_patch_size=16)  # was 14
   

2. Video metadata: New return format

   images, videos, video_kwargs = process_vision_info(
       messages,
       image_patch_size=16,
       return_video_kwargs=True,
       return_video_metadata=True  # New for Qwen3-VL
   )
   

3. Architecture improvements: Interleaved-MRoPE, DeepStack, Text-Timestamp alignment

4. New capabilities:

   • 3D grounding
   • Enhanced visual coding
   • Better spatial reasoning
   • Extended OCR (32 languages)

Compatibility: Most Qwen2.5-VL code works with minimal changes.

From Qwen2-VL to Qwen3-VL

Major Updates:

• Significantly improved performance across all tasks
• Native long context (256K vs 32K)
• Video understanding enhancements
• New architectures (MoE, Thinking editions)
• Broader model size range (2B-235B)

Future Roadmap

The following items are planned but subject to change:

• Additional model sizes and variants
• Enhanced fine-tuning tools and examples
• More cookbook examples and tutorials
• Extended language support for OCR
• Performance optimizations
• Community-contributed extensions

Resources

Documentation

• Installation Guide
• Quickstart
• Model Cards
• Benchmarks

Code & Models

• GitHub Repository
• HuggingFace Collection
• ModelScope Collection
• Cookbooks

Papers

• Qwen3-VL Technical Report (2025)
• Qwen2.5-VL Technical Report (2025)
• Qwen2-VL Paper (2024)

Community

• Discord
• WeChat Group
• GitHub Discussions

Citation

If you use Qwen3-VL in your research, please cite:

@article{Qwen3-VL,
  title={Qwen3-VL Technical Report},
  author={Qwen Team},
  journal={arXiv preprint arXiv:2511.21631},
  year={2025}
}