Original title: Axion1-0/Axion-1.0-35B
Axion 1.0 is an open-source agent foundation model recipe proposed by teams from Duke University, ByteDance, and others: first perform SFT on difficult scientific problems (learning reasoning), then perform SFT on verifiable tasks requiring information retrieval and tool use (learning investigation), and finally apply reinforcement learning on agent trajectories, with approximately 1.8 ten-thousand SFT data samples. This repository is a 35B reproduction checkpoint trained according to this recipe (total parameters approximately 350 hundred million, activated parameters approximately 30 hundred million, based on Qwen3.6-35B-A3B). The authors report scores of 52.8 on HLE, 60.4 on SEAL-0, and 85.0 on FrontierScience Olympiad problems, which are close to the 52.2, 62.2, and 86.0 reported in the paper.
Why it matters Training uses only two stages of SFT plus reinforcement learning, with data, code, and reproduction weights fully open-sourced, making it suitable for researchers studying agent training recipes for comparison
Upvotes 72 · Likes 73 · Downloads 1315
Reasons
- Targeted at scientific research scenarios: scientific reasoning, information retrieval, and tool-assisted problem solving; the model card lists scores for both the paper version and the reproduction version side by side
- The training dataset and complete training code (data generation and filtering, two-stage SFT, asynchronous reinforcement learning, evaluation) are publicly available on GitHub at bytedance/axion1_0, under the Apache-2.0 license
Usage
- Run
vllm serve Axion1-0/Axion-1.0-35B; the example uses 8 GPUs for tensor parallelism and 131072 context length, enabling--reasoning-parser qwen3and--tool-call-parser qwen3\_coder - Running agent tasks with tools requires a separate framework to execute tool calls; see the inference guide in the GitHub repository
Risks
- The GitHub description states this is a community reproduction, not an official Axion checkpoint
- Scores for other models in the comparison table are mostly taken from their respective official reports, and evaluation settings are not fully consistent
- This entry has not been run; refer to the model card for actual performance