Baseten makes Serverless Loops generally available
Baseten's Serverless Loops training SDK is now generally available, letting teams fine-tune open-weight models with SFT or reinforcement learning and ship the results to production inference with one click.
Baseten announced that Serverless Loops is now generally available. As we reported earlier, Loops is the company's SDK for supervised fine-tuning (SFT) and reinforcement learning (RL) on open-weight models. It uses a Tinker-compatible API, and Baseten manages the GPUs, kernels, model architectures and distributed training.
At launch, customers can post-train a selection of popular open-weight models at a 256k context length for both SFT and RL. Billing is per token. Trained adapters can be deployed with one click to Baseten Dedicated Inference. That inference stack serves more than 15 billion requests per day, according to the company. Weights can also be downloaded in Hugging Face Safetensors format.
Baseten says the GA release handles transient network errors and GPU failures better, starts up faster and moves weights between trainers and samplers more quickly. Together, the company says, these changes let it support more concurrent workloads. Baseten says it validates the service with end-to-end RL experiments on τ-bench tasks in the retail, airline and banking domains. Results were not included in the material reviewed.
Baseten named four early users:
- NVIDIA post-trained its Nemotron Super 3.5 VL model. Intent-classification accuracy rose from 76.8% to 94%, and the model reached 87% validation F1 on receipt extraction.
- Datadog tests training ideas for its SRE agent using multi-turn agentic RL.
- LangChain Labs trains custom models on agent traces for LangSmith Engine tasks.
- Ambience Healthcare post-trained models on clinical language.
The company says new models, quantization techniques and score centering are coming next. It also plans technical write-ups, including one on MoE LoRA inference kernels that it says are now more than 4x faster.