AI BriefingQwQ-32B: Harnessing the Power of Reinforcement Learning
Key point
Qwen released **QwQ-32B**, a model that achieves reasoning performance comparable to DeepSeek-R1 through reinforcement learning.
Details
Reinforcement learning (RL) scaling is a key technology that can dramatically improve a model's reasoning capabilities beyond existing pretraining and post-training methods.
QwQ-32B is a model with 32 billion parameters that achieves performance comparable to DeepSeek-R1, which has 671 billion parameters. This demonstrates the powerful effect of applying RL to a robust foundation model trained on vast knowledge. It also integrates agentic capabilities, enabling critical reasoning based on tool use and environmental feedback.
Model training goes through a two-stage RL scaling process:
- Stage 1 (Math and Coding): Uses outcome-based rewards. Math problems are verified with an accuracy checker, and coding is verified with a code execution server, optimizing performance based on correctness.
- Stage 2 (General Capabilities): Uses a general reward model and rule-based verifiers to strengthen overall capabilities including instruction following, human preference alignment, and agent performance.
QwQ-32B was released as open weights under the Apache 2.0 license.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.