LLM fine-tuning that speaks your domain

Q: How much training data do I need?

Enough to fill a meaningful held-out test set — typically hundreds to a few thousand quality examples for a LoRA adapter. If your data isn't there yet, we'll say so and tell you what to fix first.

Q: Which models do you fine-tune?

Open-weight models (the Llama family, Mistral, Qwen, and similar) so you own the result. We pick the smallest model that passes your eval — smaller models cost less to run.

Q: What if fine-tuning isn't the right answer?

Then we say so. If prompting or RAG clears your eval bar for less money, that's what we recommend. The assessment exists to answer this before you spend anything.

Q: Who owns the tuned weights?

You do. The adapter weights and the full training recipe are delivered to you — no lock-in, no license back to us.

Packages

Proposed pricing — these anchors are proposed, not final. Your fixed quote is confirmed before any work starts.

Package	What you get	Delivery	Price
Pilot	LoRA/QLoRA adapter on one open-weight model for one domain corpus, with a base-vs-tuned eval on your held-out set.	2 weeks	$1,500
Standard	Everything in Pilot, plus training-data curation help, quantization-aware tuning for cheaper inference, and a written eval report.	4 weeks	$3,500
Scale	Everything in Standard, plus multi-domain training, a re-tuning cadence as your data grows, and deployment into your inference setup.	6 weeks	$7,500

What's included

Eval harness on your held-out questions — base vs. tuned, scored side by side.
Trained adapter weights plus the full training recipe, documented.
Quantized, deployment-ready weights sized for your inference budget.
Handoff: how to re-run evals and when to re-tune.

Honest notes, before you spend: quality eval data is the whole game — fine-tuning on messy data teaches the model mess, and we'll tell you if your data isn't ready. Not every problem needs fine-tuning: prompting and RAG are cheaper when they work, and the free assessment finds the cheapest lever that clears your bar. You own the tuned weights — delivered to you, runnable without us.

Care plan: $99/mo — drift monitoring, periodic re-evals, and re-tuning when your data shifts. Defects against the agreed scope are covered by our 14-day defect window.

Before we start, we'll ask for

Your training and eval data — or help scoping what "good" looks like for your domain.
The behaviors you want changed, with examples of right and wrong answers.
Where the model will run — your cloud or our build; either way, you own the weights.

Frequently asked questions

How much training data do I need?

Enough to fill a meaningful held-out test set — typically hundreds to a few thousand quality examples for a LoRA adapter. If your data isn't there yet, we'll say so and tell you what to fix first.

Which models do you fine-tune?

Open-weight models (the Llama family, Mistral, Qwen, and similar) so you own the result. We pick the smallest model that passes your eval — smaller models cost less to run.

What if fine-tuning isn't the right answer?

Then we say so. If prompting or RAG clears your eval bar for less money, that's what we recommend. The assessment exists to answer this before you spend anything.

Who owns the tuned weights?

You do. The adapter weights and the full training recipe are delivered to you — no lock-in, no license back to us.

Request a quote for fine-tuning

A model that speaks your domain.

Fine-tuning, proven or not, on your data