Baseten launches Loops SDK for RL training
Baseten has released Loops, a post-training SDK in early access that runs asynchronous reinforcement learning on sequences of 131K+ tokens and deploys finished checkpoints to its inference stack in one click.
Baseten announced Loops, a training SDK for supervised finetuning (SFT) and reinforcement learning (RL) of open models. The company says Loops handles the hardware layer so researchers only write their training scripts. Harvey and OpenEvidence are named as early partners.
The release extends Baseten's move from inference into training, which we covered earlier. Loops exposes a gradient step through a few primitives: forward_backward, optim_step and sample. Baseten handles sharding, memory management and parallelism, including for frontier-size models at 131K+ sequence lengths. The SDK is Tinker-compatible, and Baseten says teams can migrate with a single import change.
Baseten highlights three features:
- Async, bounded off-policy RL. Training and sampling overlap without custom sync logic. Baseten says this addresses synchronous libraries, which pause sampling on every weight update and slow long, multi-turn agentic rollouts or train on stale data.
- Long-context RL. Support for sequences of 131K+ tokens is built in.
- Dedicated infrastructure. The company pitches this as a way to get predictable throughput and run times, unlike shared serverless platforms.
The beta supports dense and MoE models in the Qwen3.5/3.6 and Kimi families. Baseten says Nemotron, DeepSeek, GLM and MiniMax support will follow "shortly."
Baseten says it is working on online, environment-driven RL and on a Rollout Manager that would separate RL inference from the trainer and sync weights directly into inference. That follows its earlier work on faster delta weight syncs. Baseten has not disclosed pricing or a date for general availability.