Fine-tuning
Training a model on your examples — usually the wrong first move for SMBs.
Fine-tuning adapts a model's weights to your data via SFT, preference tuning or distillation, and it reliably shifts behaviour and output format rather than what the model knows. For most small businesses a better prompt plus retrieval wins on cost, turnaround and maintenance, and it costs nothing to undo. Fine-tuning earns its keep on narrow, high-volume, strict-format tasks — and it carries a real lock-in risk, as OpenAI's 2026 wind-down of its own fine-tuning platform just demonstrated.
You need this when
- You run one narrow, repetitive task at high volume with a strict, stable output format — invoice classification, ticket triage, structured extraction — and token or latency cost is visible on the invoice.
- You want to distil a large expensive model down to a small cheap one that matches it on that single task, cutting per-call cost by an order of magnitude.
- You already have 1,000+ clean, consistently labelled examples from real operations, and a well-prompted base model with retrieval has measurably plateaued below your quality bar.
You can skip it when
- Your knowledge changes weekly (prices, policies, stock, staff). Fine-tuning bakes facts into weights you cannot edit — put the documents behind retrieval instead and update them by editing a file.
- You have fewer than ~1,000 clean examples. You will mostly teach the model your labelling mistakes — spend the time on a sharper prompt with 5-10 worked examples in the context.
- You have not yet measured a prompt-plus-retrieval baseline. Build the eval set first; nine times out of ten it shows the gap is a prompting or retrieval problem, not a weights problem.
The long version — open when you want the depth
What it is — in one coffee-break
Fine-tuning means taking a model somebody else trained and continuing that training on your own examples, so the weights shift toward your task. Three variants matter commercially. Supervised fine-tuning (SFT) feeds input/output pairs — a support email in, your approved reply out. Preference tuning (DPO, RLHF) shows two candidate answers and which one your team preferred. Distillation uses a big model to generate training data for a small cheap one, so the small one imitates it on your narrow task.
The mental model that saves the most money: fine-tuning reliably changes how a model behaves — format, tone, task shape — and only unreliably changes what it knows. Google's tuning documentation lists the sweet spots as classification, fixed-format summarisation, extractive question answering and persona-consistent chat. None of those is "know our product catalogue." Facts belong in retrieval, where you update them by editing a document instead of re-running a training job.
When you actually need it (and when you don't)
For most small businesses fine-tuning is the wrong first answer, and it is worth saying so plainly before anyone spends a quarter on it. A sharper prompt, a few worked examples in the context, and decent retrieval over your documents will beat a mediocre fine-tune on cost, on turnaround, and above all on maintenance. A prompt is edited in a minute; a fine-tune is edited by rebuilding a dataset and paying for another training run.
You genuinely need it when three things line up: the task is narrow and repetitive, the output format is strict and stable, and volume is high enough that dropping to a smaller model shows up on the invoice. Distilling a large model into a small one for a single high-volume job — invoice classification, ticket triage, structured extraction — is where fine-tuning most often pays for itself. Below roughly a thousand clean, consistently labelled examples, you will mostly teach the model your labelling mistakes.
The maintenance trap is real, and 2026 proved it. OpenAI began winding down its fine-tuning platform on 7 May 2026: new organisations can no longer create jobs, and existing customers lose that ability on 6 January 2027. Tuned models keep serving only until their base model retires. Anyone who built a business process on a hosted fine-tune now has an unbudgeted migration.
How to recognize good vs bad implementations
A good implementation is honest about the baseline: it ships an evaluation harness, holds out test data you never trained on, and shows the score of a well-prompted base model next to your tuned one. It versions datasets and checkpoints, names the base model you are pinned to, and lets you leave. A bad one is a button labelled "train on your data" that returns a model with no numbers attached — you cannot tell whether anything improved, and neither can the vendor.
Three questions for the demo. "Show me the eval — what does a good prompt plus retrieval score on this task, and how much did tuning add on top?" "When you deprecate the base model, what happens to mine, and how much notice do I get?" "Can I export the adapter weights and my training set and run them elsewhere?" Vague answers to the last two are how lock-in starts.
What this costs
Compute is the cheap part. Together AI charges $0.48 per million training tokens for LoRA on models up to 16B ($0.54 full fine-tune) with a $4 minimum per job; DPO preference tuning runs $1.20 per million. A 10-million-token SFT run therefore costs about €5. OpenAI billed training at $100 per hour while it still accepted jobs. Mistral's classic fine-tuning charged a $4 job minimum plus $2 per model per month in storage. Google's Vertex pricing page lists a tuned model endpoint at 1.5x the base rate — a premium that lands on every call, not just on training.
The self-hosted path is cheaper still: RunPod rents an RTX 4090 at $0.69/hr and an A100 80GB at $1.39/hr, and a QLoRA run on an 8B open-weight model finishes in a few hours for under €10 — with the adapter weights on your own disk afterwards.
The real bill is human. Two thousand curated, reviewed examples at two minutes each is roughly 67 hours of staff time; at €40 an hour fully loaded, about €2,700 before a GPU spins up — and a fraction of that again every time the base model moves.
Where to see it scored
Only 28 of the 62 tools in the RXed audit score on this element at all — most products simply do not let you adapt the model. Top of the table: ChatGPT / OpenAI Platform (9), Krea (8.5) for image and video LoRA training, Gemini / Google (8.5) and Mistral AI (8). Bottom: Claude / Anthropic (3), with no first-party tuning, and Grok (xAI) (2). Read OpenAI's 9 with the wind-down in mind — a high score for breadth of capability is not a promise the capability will still be sold to you next year.
Flashcards
Check yourself
1. Fine-tuning is most reliable at changing which of these?
2. Your policy documents change monthly. Best approach?
3. When does OpenAI stop letting existing customers create fine-tuning jobs?
4. Roughly what does a 10M-token LoRA fine-tune of a sub-16B model cost on Together AI?
5. In a small-business fine-tune, the largest real cost is usually:
Cheat sheet
- Fine-tuning changes behaviour and format, not facts. Facts belong in retrieval.
- Prompt, then few-shot, then RAG. Fine-tune only when those measurably plateau.
- Under ~1,000 clean labelled examples, don't bother — you'll train in your errors.
- Together AI LoRA: $0.48 per 1M training tokens, $4 job minimum. Compute is cheap.
- Self-host: QLoRA on a rented RTX 4090 at $0.69/hr — an 8B run costs under €10.
- Real cost is the dataset, plus a re-run every time the base model moves under you.
Who actually does this well
| Best on this element | Score | Why it scored that |
|---|---|---|
| Cresta | 8.5 | Fine-tuning on your proprietary conversation data is the core differentiator; G2 reviewers confirm it works but requires dedicated internal investment before it performs. |
| Gemini / Google | 8.5 | Supervised fine-tuning for Gemini models on Vertex with evaluation-service integration, plus PEFT/Axolotl tuning for open models in Model Garden. |
| Krea | 8.5 | Deepest customization in the category: LoRA training on styles, characters and objects (50 images on Basic up to 2,000 on Max), Krea 2 LoRAs, and a Raw checkpoint released explicit |
| Gemini Enterprise Agent Platform | 8 | Supervised tuning, distillation, reinforcement learning and full custom training all live on the platform, with Model Registry and Model Monitoring behind them. Deep, but it is the |
| Ideogram | 8 | Real customer model adaptation on two paths: a productized Custom Models offering that trains an on-brand model from approved assets, plus open weights (Ideogram 4.0) that teams ca |
| Meta Llama ecosystem | 8 | Full-precision BF16 weights are released under Apache 2.0 explicitly for fine-tuning, and 38 finetunes plus 12 adapters are already published against Glimmer; Llama weights remain |
| Mistral AI | 8 | Hosted fine-tuning including classifier fine-tunes on the Ministral line, plus Forge for training and aligning fully custom models on a company's own knowledge — a deeper customiza |
| TensorZero | 8 | The deepest adaptation surface any gateway shipped: supervised fine-tuning, RLHF, distillation with programmatic data curation, and dynamic in-context learning, all driven by your |
And the other end of the same column:
| Weakest | Score | Why it scored that |
|---|---|---|
| Sana | 2 | No customer fine-tuning. Sana Learn personalises learning paths adaptively, which is recommendation, not model adaptation. Founded as an ML personalisation company, but none of tha |
| Workato | 2 | Workato adapts no models. BYOLLM lets you point a genie at a model you fine-tuned somewhere else, and the AI by Workato connector explicitly does not train on user data. Model adap |
| tl;dv | 2 | No customer model adaptation of any kind. Custom vocabulary and note templates are lexicon and prompt level, not tuning. |
| Avoma | 1.5 | No customer fine-tuning of any kind. Custom scorecards, templates and Smart Topics are configuration, and Avoma contractually does not train on customer data — good for privacy, bu |
Scored on 80 of 163 audited tools. Every score links to the full audit and its reasoning.