Skip to main content

ByteDance researchers detail Axion 1.0’s streamlined training

22 NOVEMBER 2026·2 MIN READ·6 SOURCES

ByteDance researchers and collaborators from three US universities released Axion 1.0, a set of AI agent models trained with two supervised fine-tuning stages followed by reinforcement learning. Their paper, submitted to arXiv on November 2, 2026, describes the approach and reports benchmark results across several tasks.

Key takeaways · 3

  • 01

    The training pipeline has two supervised fine-tuning stages followed by reinforcement learning, with no dedicated mid-training, DPO or on-policy distillation.

  • 02

    The main models are 35B and 122B, with 3 billion and 10 billion active parameters, respectively.

  • 03

    Treat benchmark comparisons as indicative rather than definitive: settings may differ across sources, and reported scores are best single-run results from multiple runs.

A shorter training recipe

The first SFT stage is designed to build self-contained reasoning using model-aware data construction.[4] The second teaches tool-dependent reasoning through automatically verifiable tasks grounded in real-world data and deterministic computation.[2] The RL stage is intended to improve long-horizon reasoning and interaction with tools and environments.[2] The Decoder reports that the paper argues lengthy nested pipelines can make it harder to identify useful steps and more expensive to reproduce.[6]

Models, data and compute

The main models are Axion 1.0-35B and Axion 1.0-122B, with 3 billion and 10 billion active parameters, respectively.[5] The project site says training used approximately 18,000 examples; the repository lists 10,819 trajectories for the first SFT stage and 7,544 for the second, totaling about 312 million tokens.[4][5] The paper reports about 1,000 GPU-hours for the 122B model, while its two SFT stages took 240 GPU-hours on 32 GPUs.[3][5] SFT runs used a 128,000-token context window.[5]

Benchmark results are mixed

The paper reports Axion 1.0-35B scoring 62.2 on SEAL-0, compared with 53.2 for both GPT-5.6 Sol and Gemini 3.1 Pro in its table.[3] Axion 1.0-122B scored 87.0 on FrontierScience Olympiad, above Gemini 3.1 Pro at 85.0 and GPT-5.6 Sol at 79.0; it also scored 66.7 on FrontierScience Research, compared with 63.3 for GPT-5.6 Sol.[1][3] On other tests, Axion 1.0-122B scored below Claude Fable 5 on full Humanity’s Last Exam and the human-difficult BioMysteryBench subset.[1]

Resources and comparison caveats

The repository includes materials for data generation and filtering, prompts, two-stage SFT, asynchronous RL, and benchmark preparation and grading; training data is available on Hugging Face.[5] The available Axion 1.0-35B Hugging Face checkpoint is identified as a community reproduction, not an official Axion checkpoint.[5] That reproduction’s weights are offered as a free download under Apache 2.0.[3]

For teams evaluating agent-model training, Axion offers a concrete recipe to examine alongside the reported results. Compare the pipeline and resources with your own requirements, and treat benchmark scores as one input rather than a direct prediction of performance in your environment.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 22 November 2026

    ByteDance researchers detail Axion 1.0’s streamlined training

Sources

AI fluency, one session a day, built for your work.