Alibaba's PAI Releases Open-Source AgenticQwen Model: 8B Version Approaches 235B Performance via Dual Data Flywheels

Gate News message, April 27 — Alibaba’s PAI team has released and open-sourced AgenticQwen, a small-scale agentic language model designed for industrial-grade tool-calling applications. The model comes in two versions: 8B and 30B-A3B. Trained through an innovative “dual data flywheel” reinforcement learning framework, AgenticQwen achieves near-trillion-parameter model-level agentic capabilities while significantly reducing inference costs.

The dual data flywheel mechanism addresses the homogenization problem in traditional synthetic data. The reasoning flywheel automatically generates harder variants from model errors, while the agentic flywheel expands simple linear workflows (such as single booking processes) into multi-branch behavior trees with constraints, rejections, and adversarial conditions, simulating real-world complex decision scenarios. Benchmarks show AgenticQwen-8B scored 47.4 on real tool environment benchmarks (TAU-2 and BFCL-V4), far exceeding the base Qwen3-8B (23.8) and approaching Qwen3-235B (52.0). AgenticQwen-30B-A3B (with only 3B parameters activated) achieved 50.2.

The model has been deployed in internal production systems similar to Manus, significantly narrowing the gap with 235B models in end-to-end inference time. However, the model is limited by a native context length of 40K tokens, which constrains its performance on deep search tasks.