arXiv paper 2611.03233 by Yuqi Jia, Yunjie Tian and five co-authors introduces Axion 1.0, an agent foundation model that is open beyond its weights. It uses just two lightweight SFT stages plus reinforcement learning, with no dedicated mid-training, DPO or on-policy distillation, and about 18K SFT trajectories in total. Training code is hosted under the bytedance GitHub organization, and the training data and an Axion 1.0-35B reproduction checkpoint (Qwen3.6-35B-A3B base, 35B total / 3B active, Apache 2.0) are on Hugging Face; weights for the paper's 122B model are listed as coming soon.
Key Takeaways
- ✓A minimal recipe: SFT stage one for self-contained scientific reasoning, stage two for tool use and information seeking, then RL for long-horizon interaction, with 18,363 SFT trajectories and 484 RL tasks
- ✓The release covers data construction, generation, filtering, verification, prompts and training code (GitHub), not just weights
- ✓The 35B reproduction checkpoint scores 52.8 on HLE, 60.4 on SEAL-0, 85.0 on FrontierScience Olympiad and 55.0 on FrontierScience Research (HF model card)
- ✓The paper's 122B model reports 59.0 on HLE with restricted website access and 66.7 on FrontierScience Research; 122B weights are not out yet
- ✓All scores are self-reported; the HF card lists Claude Opus 5 at 63.6 on HLE, so Axion still trails the strongest closed models
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and pain point
Open agent models keep multiplying, but most release only weights. Training data, synthesis pipelines and RL environments stay private, so outsiders cannot reproduce them or tell which stage produced which capability. Common recipes also stack mid-training, DPO and on-policy distillation, which adds cost and complexity. arXiv paper 2611.03233, published in early November, introduces Axion 1.0 to show that a strong agent foundation model can be both open and simple.
What is new
Axion 1.0 has three stages. The first SFT stage builds hard problems in a model-aware way, adversarially strengthening them based on the base model's own successful solutions, to train self-contained scientific reasoning. The second SFT stage builds automatically verifiable tasks from real public data sources and deterministic computation, to train tool use and information seeking. Asynchronous RL then strengthens long-horizon reasoning and tool interaction. The two SFT stages total about 18K trajectories. The 35B model is built on Qwen3.6-35B-A3B (an MoE with 35B total and 3B active parameters) and the 122B model on Qwen3.5-122B-A10B.
Numbers
All figures are self-reported. On the HF model card, the 35B reproduction checkpoint scores 52.8 on HLE, 35.3 on BioMysteryBench, 60.4 on SEAL-0, 45.9 on HiPhO, 85.0 on FrontierScience Olympiad and 55.0 on FrontierScience Research, close to the paper's 35B results. The GitHub README says the paper's 122B model scores 59.0 on full HLE with restricted website access, 87.0 on FrontierScience Olympiad and 66.7 on FrontierScience Research. The same table lists Claude Opus 5 at 63.6 on HLE, so a gap to the strongest closed models remains.
Getting started
The 35B reproduction checkpoint is Apache 2.0 and can be served with vLLM as an OpenAI-compatible endpoint per the model card (the example uses 8 GPUs, BF16, 131K context and the qwen3 reasoning and tool parsers); tool-using agent evaluation also needs your own tool-execution harness. The Axion1-0/axion-training-data dataset and the GitHub code can be used to reproduce or adapt the training pipeline. Weights for the 122B model are listed as coming soon.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.