Custom-trained n8n model.
Compact models I fine-tuned myself that turn a plain-language request into a working n8n workflow. On the task they match frontier models like Claude Opus — while running self-hosted on a single GPU, for a fraction of the API cost.
- Industry
- Own product / model fine-tuning
- Timeline
- 2025–present
- Outcome
- Frontier-level quality · ~$0.01 per workflow, self-hosted
The problem
AI that builds automations is genuinely useful — but the models good enough to do it well are expensive frontier APIs you rent by the token. At volume, or with sensitive data, that's a real constraint: the cost grows with every workflow, and your data has to leave your infrastructure for a third party. The open question was whether a small model you actually own could do the same job.
What I built
I fine-tuned my own family of compact models that turn a plain-language request into a working n8n workflow:
- trained on real work, not theory — a corpus of ~48,000 real n8n workflows, distilled down to the patterns that actually hold up in production,
- a small router decides the intent — create a new workflow or edit an existing one — then hands off to the specialist model for that job,
- the workflow comes out as one validated operation — checked against n8n before it ships, not guessed at.
The whole family runs self-hosted on a single GPU — no external API, and your data never leaves your own infrastructure.
The outcome
In my own benchmark the model matches the quality of frontier models like Claude Opus on n8n workflow generation — and it runs on a single GPU you can rent for about $0.15 an hour, which works out to roughly a cent per generated workflow: a fraction of what a frontier API would charge for the same job. Working proof that for a well-defined task, a specialized model you own outright can rival the biggest general-purpose APIs — without the per-token bill or the dependence on a single provider.