The Future of Qwen: A Generalist Agent Model
summary
Qwen3 arrives with concrete claims: a flagship 235B total with ~22B activated dense/MoE mix, a deployable 4B model for mobile, and RL fine-tuning that lifts reasoning scores from about 65 to ~80 on math/coding benchmarks. New features include hybrid thinking mode with dynamic thinking budget, 119 languages, tool-using agent behavior (function calls + environment feedback), a 7B omni prototype handling text vision audio, open-source code and checkpoints, and producible demos like webdev and deep research.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:31Important Links
- 1:34Qwen Models
- 5:20Hybrid Thinking Mode
- 11:18Models
- 12:35Multimodel Models
- 14:48Omni Models
- 15:51Open Sourcing
- 17:57Products
- 20:57Future
related talks
15mOlmo-Thinking: Training a Fully Open Reasoning Model
34mMapping the Open Model Landscape in 2025
54mGenerative to Agentic to Quantum - The Evolution of AI
27mThe Cutting Edge of AI: From LLMs to Multimodal Intelligence
36mFrom Theory To Deployment: Building Scalable Multi-Agent Question-Answering Syste...
15mAgentic Reasonings with Reinforcement Learning
47mOwn Your LLM: Running Generative and Agentic AI Locally
19mRedefining Intelligence: How Reasoning Is Re-Shaping AI in 2025