01
MODEL
Multimodality must be native.
Vision, video, documents, audio, interfaces, and space belong inside the model’s reasoning loop—not as an attachment added after the text model is built.
MULTIMODEL-NATIVE INTELLIGENCE
Shuai Bai / 白帅
Lead, Qwen-VL & Visual Intelligence · Alibaba Qwen Team
Building multimodal-native intelligence that can understand the world, use tools, and act across digital and physical environments.
01 / PERSPECTIVE
MODEL · AGENT · WORLD
RESEARCH PERSPECTIVE
My work connects foundation models, agent systems, and the environments where intelligence is ultimately useful.
Across perception, reasoning, tool use, and action, the goal is a coherent system rather than a collection of isolated capabilities.
02
AGENT
Understanding must lead to action.
A multimodal agent should perceive precisely, decide when to use tools, interact with environments, inspect the result, and recover when reality disagrees.
03
WORLD
One intelligence, many environments.
The same foundation should move across browsers, documents, software, video, 3D and CAD—and ultimately from digital action to physical intelligence.
02 / MODEL EVOLUTION
FROM NATIVE MULTIMODALITY TO COWORK
MODEL EVOLUTION
Qwen 3.5 → Qwen 3.8 Max
Perception becomes useful when it is fused with reasoning, coding, environment interaction, and long-horizon execution.
QWEN 3.5
Native Multimodal Agents
Multimodality became part of the general model itself: understand visual inputs, reason, code, and use tools in one agentic foundation.
RELEASE NOTESQWEN 3.6
Real-World Agents
The focus moved from isolated capabilities to agents that can work through longer, messier, real-world tasks.
RELEASE NOTESQWEN 3.7 PLUS
Multimodal Agent Intelligence
Stronger planning, environment interaction, and multimodal execution pushed Qwen further into the agent era.
RELEASE NOTESQWEN 3.8 MAX
Coding × Cowork × Vision
A 2.4T-scale foundation for coding and cowork—where visual context, tool use, and long-horizon execution converge.
RELEASE NOTES
03 / WORK
MODEL · SYSTEM · ENVIRONMENT
01
CORE MULTIMODAL DIRECTION
Qwen 3.5 → 3.8 Max
Native multimodal models for the agent era
Bringing visual intelligence into Qwen’s general agent foundation—connecting perception with reasoning, coding, cowork, and tool use.
02
LEAD
Qwen-VL
Qwen‑VL · Qwen2‑VL · Qwen2.5‑VL · Qwen3‑VL
A sustained program in visual understanding, grounding, documents, long video, spatial reasoning, computer use, and visual agents.
03
PHYSICAL INTELLIGENCE
Qwen-VLA
Vision · Language · Action
Extending multimodal intelligence beyond screens into action across tasks, environments, and robot embodiments.
04
UNIFIED UNDERSTANDING & GENERATION
Qwen-VLo
Understand · Generate · Edit
Unifying visual understanding and generation so a model can move naturally from interpreting the world to depicting and editing it.
05
OPEN-SOURCE AGENT SYSTEM
Qwen-MM-Plugins
Make any agent harness multimodal-native
A portable capability layer for images, video, documents, web research, long-video memory, editing, 3D, CAD, and education workflows.
SELECTED COLLABORATIONS
Collaborative & participating work
CONTRIBUTOR
Qwen-Omni Series
Qwen2.5‑Omni · Qwen3‑Omni · Qwen3.5‑Omni
End-to-end text, image, audio, and video understanding with real-time natural speech interaction.
CONTRIBUTOR · FIRST GENERATION
Qwen-Image
Native text rendering · generation · editing
An image foundation model connecting strong visual semantics with complex text rendering and precise editing.
PARTICIPATING WORK · ICML 2022
OFA
One architecture · many tasks · many modalities
An early unified sequence-to-sequence framework for vision, language, grounding, and generation.
Selected systems, not a complete publication list.
FULL PUBLICATION LIST ON GOOGLE SCHOLAR PAPERS ON HUGGING FACE
04 / OPEN SYSTEM
QWEN-MM-PLUGINS
THE CAPABILITY LAYER
A multimodal capability layer for agent systems.
Qwen-MM-Plugins turns multimodal intelligence into portable, usable capabilities across the agent tools people already work with.
EXPLORE QWEN-MM-PLUGINS
CAPABILITY MAP
08 MODULES
01
Vision + documents
→
02
Audio + video understanding
→
03
Web + image search
→
04
Long-video memory
→
05
Video editing
→
06
Blender / 3D
→
07
FreeCAD / engineering
→
08
Education agents
→
05 / JOIN
BEIJING · CHINA
WE'RE HIRING
Join the Qwen multimodal team.
We welcome exceptional interns, new graduates, experienced researchers, and collaborators working on multimodal models, agents, computer use, video, documents, 3D, world models, and VLA.
Signal over pedigree. Show what you built, what you discovered, or the question you cannot stop thinking about.
START A CONVERSATIONCONNECT
Contact
06 / VISITOR MAP
GLOBAL · APPROXIMATE · COOKIELESS
GLOBAL VISITORS
Ideas travel.
This map shows approximate daily visits by world region. One visit is counted per browser each day; only the visitor's country region is used.
VISITS SINCE AUG 2026
—
REGIONAL FOOTPRINT
CONNECTING
—North America
—South America
—Europe
—Africa
—Asia
—Oceania
COUNTRY-LEVEL LOOKUP · REGIONAL COUNTS ONLY · NO COOKIE
CC0 MAP · NATURAL EARTH / WIKIMEDIA