Athanor: Four Claude Opus 5.5 Agents Post-Train Qwen3-4B in 104 Hours

A team of four Claude Opus 5.5 agents, left alone with eight-GPU nodes and a frozen evaluation harness, lifted a Qwen3-4B base model to within a point of its vendor's own instruction-tuned release in 104 hours, with no teacher model and no human research prompts, according to a paper posted to Hugging Face titled 'Athanor: An Agent-Native System for Autonomous Post-Training Research.'

The authors (Yufa Zhou, Yichen Wang, Jiakun Fan, Ziqi Shu, Zhichao Liu, Anru R. Zhang, Neiwen Ling and Zijun Wei) report the agents moved a six-task development suite from 15.9 to 67.3, against an official instruct score of 66.5, and a held-out suite from 9.2 to 48.2 against an official 47.9. 'A team of autonomous AI agents used an open research system to independently post-train a small base model into a cross-domain generalist that matched or exceeded its official instruction-tuned version,' the abstract states.

The agents did not invent new methods. They applied supervised fine-tuning on public reasoning data, reinforcement learning 'with novel rollout modifications,' weight merges across their own checkpoints, and iterative self-distillation. What makes the result unusual is that the agents chose the sequence themselves, 'reading datasets and evaluation results' independently, 'building training data from public sources,' 'writing and modifying training algorithms' and 'tuning hyperparameters without external guidance,' per the methodology.

The authors frame this as the first known demonstration of frontier coding agents autonomously post-training a raw base into a cross-domain generalist with no teacher models or human research advice. Modern frontier agents, they argue, 'already possess sufficient engineering and research ability to execute complex autonomous post-training tasks.'

The paper leaves plenty open. It never names which tasks sit in the held-out suite, never reports total compute across four eight-GPU nodes for 104 hours, and never says whether human staff silently gated any of the agents' choices. The scores come from a single campaign on one base family; the paper does not describe a replication against a non-Qwen model.

Agent-run demos have been a thick seam this week on our side, with 345 Anthropic stories tracked over the last 90 days, landing the same day DeepMind's four Gemini 4 Argon agents cleared Minecraft from a random spawn.