# Fine-tuning

> Where a Divinci fine-tune actually trains — Vertex AI Gemini, OpenAI, Cloudflare Workers AI LoRA, AWS Bedrock, Modal and Tinker.

Fine-tuners are registered as tools like everything else, but unlike the
retrieval and audio tools they are keyed to a *base model* — the vendor decides
which of its models are tunable, and that list moves.

## Registered fine-tuners

| Backend | Base models | Credentials |
| --- | --- | --- |
| **Google Vertex AI** | Gemini (supervised tuning) | Divinci-managed |
| **OpenAI** | GPT-4o, GPT-4o-mini | BYOK |
| **Cloudflare Workers AI LoRA** | Gemma 3 12B IT, Mistral Small 3.1 24B Instruct | Divinci-managed |
| **AWS Bedrock — Claude** | Claude 3 Haiku | BYOK (AWS) |
| **AWS Bedrock — Nova** | Nova Micro, Lite, Pro, Premier | BYOK (AWS) |
| **Modal (Unsloth QLoRA)** | Gemma 4 27B on H100 | BYOK |
| **Tinker** (Thinking Machines Lab) | Managed LoRA | BYOK |

Cloudflare LoRA jobs train through **Hugging Face AutoTrain** and serve the
resulting adapter on Workers AI, so a LoRA fine-tune touches two vendors even
though it is one choice in the UI.

<Aside type="caution" title="A vendor's tunable-model list is not discoverable from its API">
Only a subset of any vendor's catalog can be tuned, the subset changes without
notice, and the publisher APIs do not expose which models those are — a retired
base model can keep answering inference calls while every tuning job against it
fails. If a job fails at submission with an unhelpful error, check the base
model is still tunable before debugging your dataset.
</Aside>

<Aside type="caution" title="BYOK-only backends refuse before they start">
On a backend that has no Divinci-managed credential, a workspace with no
provider key attached gets `FORBIDDEN` at job creation — not a queued job that
fails later. Attach the key first.
</Aside>

## What actually decides the outcome

The backend matters less than the data. A fine-tune reproduces the shape of what
you feed it, so the leverage is in
[training-data generation and flagging](/server/fine-tuning/) — the
conversation flaggers that drop turns with no responder, duplicate dates, or
bare URLs do more for quality than switching vendors.

<Aside type="tip" title="Give a fine-tuned release a fallback">
A release whose primary assistant is a fine-tune and which has no
`fallbackAssistants` has nothing to fail over to when the fine-tune errors. Set
one.
</Aside>

## See also

- [Fine-tuning & Training Data](/server/fine-tuning/) — the API and workflow.
- [Model Catalog](/guides/models/)