LangChain has launched LangSmith Fine-Tuning and a command-line tool called smithtune, linking recorded agent behavior to model training and evaluation. The public-beta workflow converts successful LangSmith trajectories—ordered records of messages, tool calls and results—into datasets for supervised fine-tuning.
The tool can pull and filter traces, help reviewers label good examples, create persistent training and evaluation splits, and check whether records fit a model’s sequence limit. Training runs through Fireworks or Baseten, so users do not need to provision GPUs. Afterward, smithtune replays held-out trajectories against the base and tuned models and uploads comparison results to LangSmith.
LangChain reports that a fine-tuned Kimi K3 model scored 96 on an internal issue-detection benchmark, compared with 90 for the base model and 87 for GPT-5.6 Sol. In a separate internal code-review test, fine-tuning Qwen-3.8-27B raised F1 from 48.9% to 53.7% while reducing model calls by 29.8%. These are company-run evaluations, not independent benchmarks.
The company cautions that an earlier, less selective dataset reduced quality. It recommends improving an agent’s harness first and using fine-tuning when repeated tasks still show consistent errors. Access requires LangSmith traces, smithtune and an account with Fireworks or Baseten.