Researchers released Athanor, an open system that gives AI agents the tools they need to run their own post‑training experiments. The agents get access to data, evaluation suites and job management, but they cannot call outside teacher models. In a 104‑hour test, four Claude Opus 5.5 agents, each on an eight‑GPU node, took the raw Qwen3‑4B‑Base model and improved it dramatically.
What the agents achieved
On a six‑task development suite the model’s score rose from 15.9 to 67.3. On a hidden held‑out suite it jumped from 9.2 to 48.2, beating the official Qwen3‑4B‑Instruct‑2507’s 66.5 and 47.9 respectively. The agents arrived at these gains by recombining known tricks: supervised fine‑tuning on public reasoning data, a new rollout change for reinforcement learning, weight merges across agents, and rounds of self‑distillation. All of this happened without any human‑provided research advice.
Why it matters
If AI agents can run reliable research loops on their own, the cost of developing better models could drop and progress may speed up. For developers who lack large research teams, an open platform like Athanor could let smaller groups experiment with autonomous training pipelines.
Why it matters
The result shows that current AI agents already have enough engineering ability to run a full post‑training research cycle. This could let organizations without big labs create competitive instruction‑tuned models more cheaply, though independent verification of the gains is still pending.
