Fine-tuning · Experiential Labs

Turn on training and your traffic becomes a model for your work. It serves what it passes on your evals and the weights are yours.
Traces from your traffic
A model trained on your work that serves what it passes. Your frontier model stays as fallback.
Claims research · August 2026
A trained Qwen3.5-4B on 500 held-out claims from the public AVeriTeC benchmark.
Computer use · July 2026
Qwen 3.5 9B distilled from Kimi K2.6 working 1,000 tasks on Mac VMs.
The claims study put a small router in front of three models. Each point is measured on 489 claims the router never saw.
Verdict accuracy against cost per 1,000 claims
Cost per 1,000 claims
95% of Opus accuracy at 31% of its cost
Enable prompt and response logging. We’ll let you know when you have enough data for a training run. You always start the run yourself.
We use your traces to simulate your production environment around the tasks most important to your agent, creating a dataset specific to its work.
We fine-tune an open small language model using reinforcement learning in the simulation.
Your fine-tuned model goes back on the gateway. Use it through Experiential like any other model, with the same API and a new model name.
A managed training engagement. If the model ships the fee goes toward its inference.
© 2026 Experiential Labs