Baseten makes Serverless Loops generally available for SFT and RL post-training
Company blog says Loops offers Tinker-compatible APIs, per-token billing, 256k context on selected open models, and one-click deploy to Baseten Dedicated Inference.
Why it matters
Baseten announced general availability of Serverless Loops on November 6, 2026: a managed training path for supervised fine-tuning and reinforcement learning on frontier open-weight models, billed per token. The company highlights a Tinker-compatible API, 256k context for the initial model set, and direct deployment of resulting adapters onto Baseten Dedicated Inference. Named design partners include NVIDIA, Datadog, LangChain Labs and Ambience Healthcare. Loops targets teams that want post-training without operating their own distributed training stack.
Key facts
- GA announced November 6, 2026 on Baseten blog
- Supports SFT and RL via Tinker-compatible API
- 256k context on initial open-model set
- One-click deploy adapters to Dedicated Inference
- Customers cited: NVIDIA, Datadog, LangChain Labs, Ambience