Fine-tune an open model on your data, and own the result
Send labeled examples and get back a model that does your task (classification, extraction, routing), checked on held-out data. The weights are yours: run them anywhere, with no per-call fees to us.
What you get
- An open model (Llama, Qwen or Mistral) fine-tuned on your examples
- An evaluation against the base model on examples it never saw
- The weights, handed over. We don't host them.
A real run on public data
- Base model: Qwen2.5-0.5B, LoRA fine-tune
- Task: Banking77, 10 intents
- Accuracy on 400 held-out examples: 60.75% base, 97.75% fine-tuned
Weights, script and evaluation: huggingface.co/booleai/qwen2.5-0.5b-banking77-10intent-lora. This is not a customer result.
Also: batch embeddings
We embedded 2,181 public-domain paragraphs with all-MiniLM-L6-v2 in 6.6 seconds, for search and clustering.
A separate fine-tune and evaluation was previously quoted at $2,000, agreed in writing before any work. The private AI sprint on this site is the flat $3,000 deployment below. There is no checkout link.
$3,000 flat
No ongoing fees. You own the deployment. Mutual NDA before kickoff.
Live within 48 hours of kickoff and data access, or a full refund.
And your files stay private
A private deployment lowers the risk of AI leaks. It doesn't replace your own security program and doesn't remove any compliance, ethics or contractual obligations (including SOC 2, if you're subject to it). We don't hold SOC 2 or any other certification. It can run fully offline on your hardware. Client data stays inside your environment and goes to no outside AI vendor. Full details: how we handle your data.
The same approach powers models you can run on your own consumer hardware. Local models for laptops, desktop GPUs, and Apple Silicon. That track has no listed price.