Axion 1.0 is an open AI agent model with a 3 step recipe

Researchers released Axion 1.0, an open model made to work as an AI agent. An agent is software that uses tools to do tasks. The authors say that it was made in only 3 steps. They say that it is competitive with leading commercial models on some tests. The summary gives no scores. The authors also release the prompts and training code.

How it works

• Step 1: supervised fine-tuning uses example answers. The group makes hard problems from methods that the first model used correctly.
• Step 2: more supervised training on tasks that use tools. Software can check each task.
• Step 3: reinforcement learning gives rewards for long tasks that use tools.

The numbers

The summary gives no scores. The authors say that Axion gives good results on many agent tests. They say that it is competitive with leading commercial models on some of them. They do not name the tests or the models.

What is not known

Only the authors make these statements. No test by other groups is given. It is not known how the model does on your tasks. The authors say that they release the data steps, prompts and training code. Other groups can then do the work again.

Insight: Open parts now exist at each layer of the agent stack

In October 2026, agent tools went to a neutral foundation. Perplexity opened its sandbox, a safe space that limits the tasks of agents. At this time, a full model recipe is open too. This reads as a push to make open parts the standard. Control can move from vendors of closed models to groups that operate open models. Leaders can have more alternatives.