Klear-AgentForge: Forging Agentic Intelligence through Posttraining Scaling

---

News

• 2025-11-12 — Released the Klear-AgentForge-8B checkpoint and published our technical report.

• 👉 Klear-AgentForge-8B

• 👉 Technical Report

• 2025-10-23 — Released the Klear-AgentForge-8B-SFT checkpoint and published the initial SFT training data. New SOTA results on SWE-bench Verifed (~8B models).

• 👉 Klear-AgentForge-8B-SFT

• 👉 Training Data

• 👉 mini-swe-agent-plus

---

Overview

Our research project Klear-AgentForge focuses on developing agentic AI capabilities through systematic post-training scaling. We will progressively update and opensource our models, data and training recipes.

Models

We present the Klear-AgentForge models, a series of agentic variants based on Qwen3, designed for complex tasks like tool use and coding. Currently, we provide two models:

• Klear-AgentForge-8B-SFT: Trained from the Qwen3-8B base model via Supervised Fine-Tuning (SFT) on a mixture of data from various agentic tasks.
• Klear-AgentForge-8B: Further trained from the Klear-AgentForge-8B-SFT model using reinforcement learning, followed by a model merge.

Evaluation results

Tool Use Performance

Model	BFCL v3	Tau Bench Retail	Tau Bench Airline
Qwen3-30A3B-2507-Instruct	65.1	59.1	40.0
Qwen3-30A3B-2507-Thinking	72.4	67.8	48.0
Qwen3-8B (Non-Thinking)	60.2*	35.7*	12.0*
Qwen3-8B (Thinking)	68.1*	45.2	25.0
xLAM-2-8B-fc-r	72.8	58.2	35.2
AgentScaler-8B	—	50.2	42.0
Klear-AgentForge-8B	71.5	56.7	41.5

Note: Results with * are our reproductions; all other baseline results are from official reports.

Coding Performance

Model	SWE-bench Verified	Aider Polyglot
Qwen3-30A3B-2507-Instruct	17.6*	35.6
Qwen3-8B (Non-Thinking)	8.0*	16.4*
Qwen3-8B (Thinking)	9.8*	17.3*
SWE-Mirror-LM-7B	22.8	—
SWE-agent-LM-7B	15.2	—
Klear-AgentForge-8B	39.4	33.8

Note: Results with * are our reproductions; all other baseline results are from official reports. SWE-bench Verified was evaluated with mini-swe-agent-plus.

SFT training data

SWE-smith-mini_swe_agent_plus-trajectories-66k

Contact

For questions or collaborations, please open an issue in this repository.

Citation

If you find this project is useful in your own work, please consider citing as follows:

@misc{klear_codetest,
    title = {Klear-AgentForge: Forging Agentic Intelligence through Posttraining Scaling},
    url = {https://github.com/Kwai-Klear/Klear-AgentForge},
    author = {{Qi Wang, Hongzhi Zhang, Jia Fu, Kai Fu, Yahui Liu, Tinghai Zhang, Chenxi Sun, Gangwei Jiang, Jingyi Tang, Xingguang Ji, Yang Yue, Jingyuan Zhang, Fuzheng Zhang, Kun Gai, Guorui Zhou}},
    month = {November},
    year = {2025}
}

---

Stay tuned for updates!

⭐ Star this repository to follow our progress