Axion 1.0 is an open agent model trained in three stages, and the authors say the whole recipe is public.
Training: Two lightweight supervised fine-tuning stages, then reinforcement learning. The paper skips mid-training, DPO and on-policy distillation. The submitter says it used about 18K SFT examples.
Release: The paper releases the data construction, filtering, verification, prompts and training code. The training data and code are linked from the page.
Model: The submitter says the 35B reproduction checkpoint is available under Apache 2.0, with 52.8 on HLE and 85.0 on FrontierScience Olympiad. The 122B model is mentioned, but the page only links the 35B checkpoint.
Benchmarks: The abstract says it is competitive with leading commercial models on several agentic benchmarks.
Sign in to comment.
TL;DR of the paper, generated automatically.
Axion 1.0 is an open agent model trained in three stages, and the authors say the whole recipe is public.