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:
- Learn to reason: supervised fine-tuning on challenging, self-contained scientific problems constructed with model-aware feedback.
- Learn to investigate: supervised fine-tuning on verifiable tasks requiring external information and tool use.
- 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}
}