Arctic-MidTool-RL-4B
The final 4B agent from the MidTool recipe: Qwen3-4B-Base → mid-training on MidTool-Mix → tool-use SFT → agentic RL. Start here if you want a model to run, rather than one to train further (Arctic-MidTool-MT-4B is the mid-training checkpoint).
Results
| 4B model | BFCLv3 Overall | τ²-Bench Pass@1 | MCP-Universe Score |
|---|---|---|---|
| Qwen3-4B (official) | 24.27 | 11.87 | 16.05 |
| Qwen3-4B-Base + SFT + RL | 39.51 | 13.04 | 17.65 |
| Arctic-MidTool-RL-4B | 54.18 | 19.96 | 19.73 |
Largest gains are on multi-turn BFCL (27.63 vs 19.00) and τ²-Bench retail/airline — the settings that need long-horizon execution and recovery.
Tool calls are emitted in the Qwen3 <tool_call> format; pass your schemas via the tools= argument of the chat template. Thinking mode is disabled in our evaluation setup.
Details
Mid-trained on MidTool-Mix; see the dataset's LICENSE for data terms.
See our paper for the full data, training, and evaluation details.
@article{jiang2026midtool,
title = {MidTool: Mid-training Data Synthesis for Agentic Tool Use},
author = {Jiang, Fengqing and Wang, Yite and Liu, Boyi and Wang, Zhaoyang and
Xu, Canwen and Yao, Zhewei and Poovendran, Radha and He, Yuxiong},
year = {2026}
}