RL and post-training loops for models and agents. Evolutionary search and robots are coming soon. Pick a policy, an environment and an algorithm. Spacecraft runs the loop on your GPUs and shows you what changed.
Models and agents today, robots soon. Sandboxes, graders, GPU jobs and logs, set up for you.
SFT, GRPO or your own loop, on your own GPUs. Weights land in your Hugging Face repo. Evolutionary search is coming soon.
hf:// exportWatch loss and reward live. Compare the base, any checkpoint and the final policy side by side.
Every step is logged. Jump back to step 12, fork it and compare the rewards.
The policy acts in one sandbox. The grader scores it in another. The log sits where neither can write.
Science moves at the speed of its experiments. Spacecraft makes every RL and evolutionary loop fast to start, safe to trust and easy to repeat, so discovery can run on its own.
No. Connect Modal and Hugging Face once. Pick a policy, an environment and an algorithm. We handle the rest.
Agents, yes. If it acts in an environment and gets a score, Spacecraft runs the RL loop. Robots and evolutionary search are coming soon.
No. Runs use your Modal account. Your GPUs, your bill.
OpenEnv on Hugging Face, verifiers on the Prime Intellect hub, or any Hugging Face dataset.
Yes, with TRL. GRPO from a Colab notebook works today.
Yes. Audit checks for reward hacks, answer leaks and flaky rewards before you train.
Soon. Dedicated inference is on the way.
Gemma 4 E2B was trained with RL on Spacecraft for $0.59 of compute.