LangChain's applied research team trains the models behind LangSmith Engine on Baseten Loops. They fine-tune an open model on agent traces from LangSmith so it can diagnose failures in a user's connected GitHub repo, and we wrote about that work earlier this month.
That workflow is now open to anyone building on LangSmith. LangChain announced LangSmith Fine-Tuning today at Interrupt in New York. It runs on smithtune, an open-source CLI you can use yourself or hand to a coding agent.
The CLI reads your LangSmith traces and turns the successful runs into a supervised fine-tuning dataset. Training then runs on Baseten Loops. Once LangSmith has evaluated the new checkpoint, smithtune deploy puts it on a Baseten Dedicated Inference deployment.
Training on Baseten
With Baseten as the training provider, smithtune trains through the Loops Python SDK on dedicated GPUs in your own Baseten workspace. Checkpoints are saved on Baseten during the run, so deploying one never means downloading weights and uploading them somewhere else.
Before scheduling a run, the CLI checks which models your workspace is enabled to train. This mostly helps when a coding agent is typing the commands. It can see what's available and won't queue a job that can't start.
You use your own Baseten API key, and training and inference bill to your Baseten account.
Get started
The code is at github.com/langchain-ai/smithtune. Training on Baseten needs an account with Loops enabled, and because Loops is in early access, you may need to request it for your workspace first.
If a coding agent will be running smithtune, the agent skill in the repo points to docs.baseten.co/agent-setup. That page sets your agent up with our agent skills and the Baseten docs MCP server.
