# Axion 1.0 trains an agent model with two SFT stages and RL

> The open recipe skips mid-training and DPO, and its 122B model scores 59.0 on Humanity's Last Exam. Weights for the 122B model are not clearly released.

URL: https://claudegpt.si/news/axion-1-0-open-agent-model-recipe  
Published: 2026-11-05T02:53:56.930Z  
Author: ClaudeGPT Si Newsroom  
Section: Research  
Topics: Axion, ByteDance, AI agents, Open-weight models, Reinforcement learning  
Publisher: ClaudeGPT Si

## Key takeaways

- Axion 1.0 trains an agent model with two supervised fine-tuning stages and reinforcement learning, with no mid-training, DPO or on-policy distillation.
- The 122B model scores 59.0 on Humanity's Last Exam, below the 63.6 and 63.8 the paper lists for Claude Opus 5 and Claude Fable 5.
- Code and training data are public under Apache-2.0, but only a 35B model is on Hugging Face, labeled a community reproduction.

Axion 1.0 is an open agent model from researchers at ByteDance, Duke, Maryland and Buffalo, described in a [paper posted on arXiv](https://arxiv.org/abs/2611.03233) on November 3, 2026. Its pitch is a short recipe: two rounds of supervised fine-tuning, then reinforcement learning, with no mid-training, no preference optimization (DPO) and no distillation from a live teacher. The authors say the result is competitive with leading commercial models on several agent benchmarks, and they publish the data pipeline and training code along with the weights.

Agent models, meaning language models that search the web and run code to answer hard questions, are usually built with long, layered pipelines. When a lab releases only the final weights, nobody can tell which steps mattered. Axion is an attempt to show that a much shorter pipeline gets most of the way there, and to let others check.

## What the team did

Axion starts from existing open models and does not train from scratch. The 35B version starts from Qwen3.6-35B-A3B and the 122B version from Qwen3.5-122B-A10B. Both are mixture-of-experts models, which run only part of their network for each token (3B and 10B active parameters). The paper also reports 4B and 9B versions trained with fine-tuning only, to test how the recipe scales down.

The work is mostly in the data, in two stages.

- **Stage I, self-contained reasoning.** Starting from public course material, competition problems and reasoning datasets in math, physics, chemistry, biology and computer science, a generator model writes new questions. If the base model already solves one, the pipeline makes it harder: it adds a dependent step, swaps fixed numbers for symbols, shifts the range so shortcuts fail, or adds cases. A reference answer is kept only when GPT-5.5 gives the same answer in three attempts. Trajectories that solve the problem with Python are filtered for correctness and format.
- **Stage II, tool-dependent reasoning.** Questions are built from structured public data (biology, chemistry, geography, aviation) by filtering records, picking one entity, computing a value in code, and having a model phrase the question. The source database is hidden from the solver, and questions that can be answered without tools are dropped. The model has to search, open pages or compute.
- **Stage III, reinforcement learning.** The team uses GRPO, a reinforcement learning method that compares several sampled answers to the same question. A correct final answer scores 1. Otherwise the model gets limited partial credit for verified intermediate steps.

The project page says the training set is about 18,000 examples. The paper reports roughly 90 GPU-hours of fine-tuning for the 35B model and 240 for the 122B model, on 32 GPUs.

## What the numbers show

The authors report results on six benchmarks. Their numbers are the best of several runs, and some baseline scores come from other groups' evaluations, so the authors call the comparisons indicative rather than strictly controlled.

| Benchmark | Axion 1.0-35B | Axion 1.0-122B |
| --- | --- | --- |
| Humanity's Last Exam (HLE) | 52.2 | 59.0 |
| BioMysteryBench (hard subset) | 35.3 | 41.7 |
| SEAL-0 | 62.2 | 58.6 |
| HiPhO | 45.7 | 46.2 |
| FrontierScience Olympiad | 86.0 | 87.0 |
| FrontierScience Research | 56.7 | 66.7 |

Against the open agent model Agents-A1, the 35B model is ahead by 4.6 points on HLE, 5.8 on SEAL-0 and 7.0 on FrontierScience Olympiad. On FrontierScience Research the 122B model scores 3.4 points above GPT-5.6 Sol, according to the paper. The paper's table also lists HLE scores of 63.6 for Claude Opus 5 and 63.8 for Claude Fable 5, both above Axion's 59.0. It reports BioMysteryBench scores of 49.4 and 46.1 for the same two models, again above Axion. So the claim of competitive results holds on some benchmarks and not on the hardest one.

The stage-by-stage table is the more useful part. On a 130-question text-only HLE subset, the 122B model goes from 40.8 for the base model to 68.5 after the first fine-tuning stage and 74.6 after RL. Most of the gain on science questions comes from Stage I. Stage II mainly helps the tool-heavy benchmarks. RL adds on HLE and SEAL-0.

## Where it is weaker

The paper is open about several limits.

- HLE depends on whether the model may browse certain websites. The 35B model scores 52.2 with restrictions and 59.1 without them.
- RL does not help everywhere. For the 35B model it lowered FrontierScience Research (58.3 to 56.7) and HiPhO (46.8 to 45.7). Stage I lowered its BioMysteryBench score from 23.5 to 17.6.
- The three-stage results are sequential checkpoints, not ablations that remove one part at a time, so they do not prove each stage is needed.
- Comparing the 35B and 122B models is not a controlled scaling study, and the 4B and 9B models have no RL stage.
- The pipeline still leans on other models: GPT-5.5 verifies answers, and Qwen-based judges are used for rewards and summaries.

## What is released

The code is at [bytedance/axion1_0](https://github.com/bytedance/axion1_0) under the Apache-2.0 license. It has the data generation scripts, supervised fine-tuning and RL training code, and evaluation tools for FrontierScience, HiPhO, HLE-130 and SEAL-0. The training dataset is on Hugging Face as Axion1-0/axion-training-data.

Weights are less clear. A 35B model is on Hugging Face as [Axion1-0/Axion-1.0-35B](https://huggingface.co/Axion1-0/Axion-1.0-35B) under Apache-2.0, and the project page links it. The GitHub repository calls it a community reproduction rather than an official checkpoint, and its model card is tagged "reproduction." The repository does not link downloadable 122B weights, and it does not say whether the 4B and 9B models are out. Anyone who wants to use the 122B model should check for a release before planning around it.

## What to watch

The useful test is reproduction. The recipe is cheap by agent-model standards, so independent groups can rerun it on other base models, and report averages instead of best runs. If the 35B reproduction matches the paper's table, that supports the argument that a short, data-heavy pipeline is enough for many agent tasks. The benchmark gap to the top closed models on HLE would remain.

## Frequently asked questions

### What is Axion 1.0?

Axion 1.0 is an open agent model recipe from researchers at ByteDance, Duke, Maryland and Buffalo. It fine-tunes existing Qwen models in two supervised stages and then applies reinforcement learning.

### Are the Axion 1.0 weights available?

A 35B model is on Hugging Face under Apache-2.0, but the GitHub repository describes it as a community reproduction, not an official checkpoint. No downloadable 122B weights are linked.

### How much compute does Axion 1.0 training need?

The paper reports about 90 GPU-hours of supervised fine-tuning for the 35B model and 240 for the 122B model, on 32 GPUs. Reinforcement learning is extra.

## Sources

- [Axion 1.0: Open Agent Foundation Models, Simplified](https://arxiv.org/abs/2611.03233), arXiv
- [Axion 1.0 project page](https://axion1-0.github.io/), Axion 1.0 authors
- [bytedance/axion1_0](https://github.com/bytedance/axion1_0), GitHub
- [Axion1-0/Axion-1.0-35B](https://huggingface.co/Axion1-0/Axion-1.0-35B), Hugging Face
- [Axion 1.0 on Hugging Face Papers](https://huggingface.co/papers/2611.03233), Hugging Face