Tinker-Atropos is a 📝Nous Research integration layer connecting the Atropos reinforcement learning environment framework to Thinking Machines Lab's Tinker training API, letting Atropos environments train models through Tinker.
Nous introduced it in November 2025, describing Atropos as a decoupled environment service that separates trainer management, rollout collection and environment setup, and Tinker as a trainer-as-a-service that fits that design. It exposes OpenAI-compatible inference endpoints that talk to Tinker's sampling client, which lets any Atropos environment built on the ManagedServer pattern run against Tinker without modification.
A training run uses three processes: the Atropos API, the trainer and an environment. Configuration follows the Atropos format with an added tinker section for settings such as LoRA rank and learning rate. The bundled example trains Llama-3.1-8B-Instruct on GSM8k for 50 steps, and the finished weights are returned as a Tinker path for download. Runs require a Tinker API key, and costs follow Tinker's published rate card.
In Nous's own comparison with Tinker's example training loop on GSM8k, the integration cost about $0.06 more in tokens, had a lower average training time per step, and reached a 0.9 reward plateau in two minutes against three. Nous suggests the speed likely comes from running inference, environments and training asynchronously.
Release notes for 📝Hermes Agent v0.3.0 (March 17, 2026) record making Tinker-Atropos RL training fully optional; as of November 2026, the Hermes docs describe the agent's RL training pipeline as powered by Atropos, whose repository is now archived.