LangSmith Fine-Tuning launches on Baseten Loops
LangChain's new LangSmith Fine-Tuning beta lets developers turn successful agent traces into fine-tuned open models, with training and inference running on Baseten under their own accounts.
LangChain launched LangSmith Fine-Tuning in public beta at its Interrupt 26 conference in New York. Training runs on Baseten Loops, Baseten said in a blog post.
The feature is built on smithtune, an open-source CLI. Developers can run it themselves or give it to a coding agent. Smithtune reads a user's LangSmith traces and turns the successful runs into a supervised fine-tuning dataset. It then trains an open model on Baseten Loops. After LangSmith has evaluated the new checkpoint, the smithtune deploy command puts it on a Baseten Dedicated Inference deployment.
Baseten says this is the workflow LangChain's applied research team already uses to train the models behind LangSmith Engine. That team fine-tunes an open model on agent traces from LangSmith.
With Baseten as the provider, smithtune trains through the Loops Python SDK on dedicated GPUs in the user's own Baseten workspace. Checkpoints are saved on Baseten during the run, so deploying one doesn't require downloading weights and uploading them elsewhere.
Before it schedules a run, the CLI checks which models the workspace is allowed to train. Baseten says this mainly helps when a coding agent is entering the commands, because the agent can see what is available and won't queue a job that can't start.
Users supply their own Baseten API key, and both training and inference are billed to their Baseten account. The post doesn't name the supported base models, give pricing, or say whether other training providers are planned.