Alibaba PAI Open-Sources AgenticQwen: "Dual Data Flywheels" Propel 8B Model to Rival 235B Performanc
Alibaba’s Platform for AI (PAI) team has officially released and open-sourced AgenticQwen, a new series of small language models specifically engineered for industrial-grade tool usage. Available in 8B and 30B-A3B parameter s<!--autointro-->
Traditional synthetic data methods often suffer from homogeneity, causing model performance to plateau. AgenticQwen overcomes this limitation through a dynamic trainingapproach involving two distinct flywheels:
Reasoning Flywheel: This mechanism automatically generates harder variants of problems based on the model's previous errors, forcing continuous improvement in logic and reasoning.
Agent Flywheel: Instead of simple linearworkflows (e.g., a straightforward booking process), this flywheel expands execution trajectories into complexbehavior trees. It simulates real-world decision-making by introducing constraints, rejection scenarios, and adversarial conditions.
Evaluations indicate that AgenticQwen delivers exceptional results on real-world tool-use benchmarks, such as TAU-2 and BFCL-V4:
AgenticQwen-8B: Achieved an average score of 47.4, vastly outperforming the base Qwen3-8B (23.8) and closely approaching the performance of the massive Qwen3-235B (52.0).
The model has already been deployed in internal production systems similar to Manus, demonstrating a significant reduction in end-to-end inference time compared to larger models. However, the team notes that due to a native context window of 40K tokens, small models like AgenticQwen still face limitations in deep searchtasks requiring extensive context retention.
Comments & Questions (0)
No comments yet
Be the first to comment!