Own your AI model. Stop paying per token.
DistillForge turns one repetitive LLM task into a small, private model your company owns and runs on its own hardware. Big-model quality on that task, small-model cost.
Who it's for
Teams with big API bills for narrow, repeated tasks: support-ticket triage, data extraction, classification, routing and short summaries.
Companies that want prompts and customer data kept inside their own network, with faster answers than a hosted large model.
What you get in about two weeks
A distilled private model
A large teacher model answers your prompts; a small student is fine-tuned on those answers with LoRA on a single consumer GPU.
A large teacher model answers your prompts; a small student is fine-tuned on those answers with LoRA on a single consumer GPU.
An honest evaluation report
The student against the model you use today, on held-out prompts: quality, speed and size, with exact-match scores.
The student against the model you use today, on held-out prompts: quality, speed and size, with exact-match scores.
The weights, delivered
A GGUF file ready for Ollama on your own machines. You own it, and there's no per-token bill after export.
A GGUF file ready for Ollama on your own machines. You own it, and there's no per-token bill after export.
Introductory pilot pricing (CAD, plus GST)
Starter pilot
about $2,500
One task, one student model, held-out evaluation report, GGUF/Ollama model delivered.Full pilot
about $5,000
Everything in Starter, plus a head-to-head with your current model on quality, speed and cost, one retraining round after your feedback, and help running it on your hardware.Introductory prices for early pilot customers. Nothing here is a committed quote until agreed in writing.
Proof from our MVP demo (measured Oct 21, 2026)
87%agreement with the teacher (base model 38%)
9 min 34 sto fine-tune on one RTX 3060
2.1x fasteranswers than the teacher (0.68 s)
9.8x smaller507 MB vs 4.9 GB
Task: support-ticket triage into strict JSON, teacher Qwen3-8B, student Qwen2.5-0.5B-Instruct, 112 training and 20 held-out prompts. Exact label match was 71% (base model 21%). This is a small sample and agreement was judged by the teacher model; a pilot uses your prompts and a larger held-out set.
What we need from you
- Sample prompts for one task (real or anonymised) and what a good answer looks like.
- Your choice of teacher: an OpenAI-compatible API or a local open-weight model. Confirm your provider allows training on its outputs; open-weight teachers such as Qwen3 (Apache 2.0) avoid that question.
- Data handling agreed in writing first. Training can run on your own hardware if you prefer.
Send one example prompt and your volume
We reply with a pilot plan. Daniel James Bertrand, Founder · 250-268-8780