Introducing LangSmith Fine-Tuning

LangChain announced LangSmith Fine-Tuning and smithtune, a command-line interface for turning LangSmith agent trajectories into custom fine-tuned models. The launch, published September 24, 2026, is in Public Beta. The CLI covers dataset creation and preparation from LangSmith trajectories, training with Fireworks or Baseten, and evaluation with LangSmith, and can be run directly or driven by a coding agent.

smithtune is built for post-training and currently supports supervised fine-tuning (SFT), which trains a model on examples of good behavior by updating model weights to imitate them. LangSmith's trajectory format records what a model saw at every turn, which the company says is intended to support SFT and to preserve tool availability context that a naive export of the final message list would lose.

The workflow involves pulling trajectories from a LangSmith tracing project with optional filters, labeling traces to identify good candidates, and storing a persistent dataset for training and evaluation. It also filters traces beyond a given sequence length, splits data into train/validation/test sets, and offers a plan step to review settings such as model, training example count, and hyperparameters. Training jobs are submitted to Fireworks managed SFT or Baseten Loops, which support LoRA training; after training, an evaluate command compares the selected checkpoint with the base model using replay evaluation, and a deploy command serves the tuned model.

LangChain said it applied the flow to two internal agents. For Engine, fine-tuning base Kimi K3 on curated trajectories scored 96.0 on a subset of IssueBench, versus 90.0 for base Kimi K3 and 87.0 for GPT-5.6 Sol. For OpenSWE Review, SFT on Qwen-3.8-27B raised F1 from 48.9% to 53.7%, with 29.8% fewer model calls and 29.4% fewer tool requests. LangChain also said an earlier, less selective training set reduced F1 after SFT.

Getting started requires a LangSmith account with agent traces, an API key for Fireworks or Baseten, and the smithtune CLI. LangChain partnered with Fireworks and Baseten on the integration.