---

license: apache-2.0 library_name: transformers pipeline_tag: text-generation datasets:

• Axion1-0/axion-training-data tags:
• reproduction
• agent
• tool-use
• scientific-reasoning
• moe
• arxiv:2611.03233

---

Axion 1.0-35B — Reproduction

Paper · Project · Code · Training Data

This repository provides a reproduction checkpoint of Axion 1.0-35B, following the recipe described in Axion 1.0: Open Agent Foundation Models, Simplified. The model is designed for scientific reasoning, information seeking, and tool-augmented problem solving.

Model Overview

Property	Details
Model	Axion 1.0-35B
Release	Reproduction checkpoint
Architecture	Mixture of Experts (MoE)
Parameters	Approximately 35B total, 3B active
Weight precision	BF16

The training recipe combines approximately 18K SFT examples with three stages:

1. Learn to reason: supervised fine-tuning on challenging, self-contained scientific problems constructed with model-aware feedback.
2. Learn to investigate: supervised fine-tuning on verifiable tasks requiring external information and tool use.
3. Learn through interaction: reinforcement learning over agent trajectories to refine reasoning and interaction with tools.

Performance

The table compares paper-reported results with the results of the reproduction checkpoint released in this repository. All scores are percentages; higher is better.

Model	HLE	BMB	SEAL‑0	HiPhO	FS Olympiad	FS Research
Agents-A1	47.6	—	56.4	46.4	79.0	40.0
Kimi-K2.6	54.0	—	50.5	41.1	73.0	17.9
DeepSeek-V4-Pro (Preview)	48.2	—	55.0	38.7	76.0	13.3
Apodex 1.1 (ReAct)	53.2	23.5	—	—	—	55.0
Apodex 1.1 (Agent Team)	56.1	35.3	—	—	—	63.3
GPT-5.5	52.2	—	42.3	43.3	78.0	26.7
GPT-5.6 Sol	—	47.1	53.2	42.9	79.0	63.3
GPT-6 Astra	57.2	—	—	—	—	—
Muse-Spark-1.0	58.0	—	—	—	—	38.0
Gemini 3.1 Pro	51.4	35.3	53.2	38.5	85.0	26.7
Claude Opus 5	63.6	49.4	—	—	—	—
Claude Fable 5	63.8	46.1	—	—	—	—
Axion 1.0-35B (Paper)	52.2	35.3	62.2	45.7	86.0	56.7
Axion 1.0-35B (Reproduction)	52.8	35.3	60.4	45.9	85.0	55.0

HLE denotes the full Humanity's Last Exam benchmark; BMB uses the human-difficult BioMysteryBench subset. FS denotes FrontierScience. A dash indicates an unreported result. Bold highlights the reproduction checkpoint released in this repository.

Usage

Following the repository's inference guide, serve the released checkpoint with vLLM:

vllm serve Axion1-0/Axion-1.0-35B \
  --served-model-name axion-35b \
  --host 127.0.0.1 --port 8000 \
  --tensor-parallel-size 8 --dtype bfloat16 \
  --max-model-len 131072 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

This example uses eight GPUs. Adjust tensor parallelism and context capacity to fit your GPU memory.

Call the OpenAI-compatible endpoint:

curl --fail-with-body http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "axion-35b",
    "messages": [
      {"role": "user", "content": "What is 2 + 3?"}
    ],
    "max_tokens": 1024
  }'

Tool-using agent evaluation additionally requires a harness that executes tool calls. See the inference guide for details.

License

This model is released under the Apache License 2.0. See LICENSE for details.

Citation

@misc{jia2026axion,
  title         = {Axion 1.0: Open Agent Foundation Models, Simplified},
  author        = {Yuqi Jia and Yunjie Tian and Jiayu Zheng and Yijun Liang and Jingchen Sun and Mengyu Zhao and Di Fu},
  year          = {2026},
  eprint        = {2611.03233},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2611.03233}
}