AI Model Fine-Tuning Services
For teams whose language, vision or speech model keeps missing their domain. We work out with you whether training is worth it, then fine-tune on your examples and test it against your current model on your own test set before you switch.
Free · You talk to an engineer, not a salesperson
Every engagement starts with a written go/no-go and a baseline score on your task. No training spend until you approve it.
We build and hand over the model. Running it in production is a separate engagement.
What are AI model fine-tuning services?
AI model fine-tuning services take a pretrained model and train it further on your examples, so it learns a behaviour the general model doesn't have: your output format, your product terms, the defects on your production line, or your callers' accents. At Crescent AI we fine-tune three kinds of model: language models (text in, text out), vision and vision-language models (images and documents), and speech models (audio to text). Before any of that, we test whether a better prompt, retrieval, or your provider's built-in customization would solve the problem for less. It's one of three methods in our custom AI model development work.
When prompting stops being enough
Language models
- “It keeps getting our formats and terms wrong.” Output that has to follow a schema, a policy or a house style drifts across thousands of calls.
- “Sensitive text can't leave our environment,” so a hosted frontier model isn't an option.
Vision and vision-language models
- “It doesn't recognise our defects.” Off-the-shelf detectors are trained on everyday objects, not your scratches, cracks or misprints.
- “It misreads our documents.” Your forms, invoices or drawings have layouts a general model wasn't trained on.
Speech models
- “It mangles our product names.” Brand names, part numbers and jargon come out as similar-sounding words.
- “It struggles with our callers' accents.” A model that does well on broadcast English does worse on regional dialects.
Fine-tune, RAG, or neither?
Fine-tuning changes how a model behaves. Retrieval (RAG) gives it your facts at answer time. Prompting and built-in customization change what you ask it, or what it listens for. We start with the cheapest one that could work.
| Your situation | Best first move | Fine-tune when… |
|---|---|---|
| The facts it needs change often or live in your documents | Retrieval (RAG) | Facts are found but answers still come out in the wrong form |
| Format, tone or policy drifts even with a good prompt | Fine-tune a language model | This is the core case |
| A speech model misses your product names | Your provider's custom vocabulary or hotwords | The list is long, terms are ambiguous, or accents are the real problem |
| A vision model misses your defect types | A zero-shot model or a managed vision API | It still misses the defects you care about, or images can't leave your site |
| You have fewer than about 100 good examples | Better prompting, or a labelling plan | You have enough examples to learn from |
Every engagement starts with a written go/no-go, and nothing trains until you sign it. If the answer is “don't fine-tune”, you keep the evaluation suite and baseline scores.
Not sure which row you're in?
Book a free fit call(opens scheduling widget)Bring 10 examples, images or audio clips the current model gets wrong.
How we fine-tune
| Method | What it is | Use it when |
|---|---|---|
| LoRA / QLoRA | Training a small set of added weights instead of the whole model (QLoRA does it on a compressed copy). The result is a small adapter file | Our default for language, vision-language and speech models: cheaper runs, easy to swap or roll back |
| Full fine-tuning | Updating all of the model's weights | Adapters plateau below your bar and the gain justifies the compute |
| Preference tuning (DPO) | Training on pairs of better and worse answers | “Correct” is a judgement call: tone, helpfulness, policy edge cases |
| Task-head fine-tuning | Retraining a vision model's final layers, which produce boxes, masks or labels, on your labelled images | Detection, segmentation and classification |
We try two or three base models where licences allow, because the best starting model differs by task.
What we fine-tune
Language models
- When: the model must hold a format, policy, tone or vocabulary that prompting can't hold reliably: a fixed JSON schema for claims, a support style, your product taxonomy.
- What we measure: task accuracy on a held-out test set, plus regression checks for instruction following, refusals and tool schemas.
- Data: hundreds to low thousands of good examples for a narrow task. Quality matters more than volume.
- Research: across 31 tasks, 4-bit LoRA fine-tuned models outperformed GPT-4 by 10 points on average (LoRA Land, arXiv 2405.00732). That is a vendor-authored average, and GPT-4 still led on broad tasks.
Vision and vision-language models
- When: you need to detect, segment or classify things a general model doesn't know (defects, parts, crop disease), or read documents and forms with your own layouts.
- What we measure: for detection, how many of each defect type the model finds and how many false alarms it raises (recall, precision, mAP); for documents, field-level accuracy against a hand-checked set.
- Data: labelled images, or a labelling plan we run with your team, plus a sample of real production images, including the awkward ones: glare, blur, rare defects.
- Managed option: Microsoft Foundry offers vision fine-tuning for selected GPT models (GPT-4o and GPT-4.1 in its October 2026 documentation) and skips images containing faces or people from training. If your task involves people, or images can't leave your site, an open-weight model in your environment is the route.
- Licences: fine-tuned Ultralytics YOLO models used in a commercial or private company setting need an Ultralytics Enterprise Licence unless the whole project is open-sourced under AGPL-3.0. We flag this in the fit memo and compare alternatives with permissive licences.
- What we don't claim: we found no published before-and-after benchmark for fine-tuning vision-language models on invoices and forms. We measure it on your documents in the Fit Sprint instead of quoting a number.
Speech models
- When: transcription keeps missing product names, jargon or your callers' accents, and the provider's custom vocabulary hasn't fixed it.
- What we measure: word error rate (the share of words transcribed wrongly), overall and on your key terms, on recordings the model hasn't trained on.
- Data: in-domain audio with accurate transcripts, and consent to use the recordings for training.
- Managed option first: Microsoft's documentation says its base speech model “works well in most speech recognition scenarios”, with custom models for domain vocabulary or specific audio conditions. Google offers phrase hints and boosts. We test these first.
- Research: in one study, fine-tuning Whisper cut word error rate on a range of British dialects from 16.8% to 1.0% (IWSDS 2025). Whisper is MIT-licensed, so a fine-tuned version can be used commercially.
Bring a few audio clips it transcribes badly, or 20 images it gets wrong.
Book a free fit call(opens scheduling widget)Need something else? A smaller, cheaper model for a task a large one already does well → model distillation. Search or RAG finds the wrong passages → embedding model fine-tuning. Your vendor is retiring your fine-tuned model → fine-tuned model migration.
What changes when it works
Each outcome is measured against your current model on the same tests, before you switch:
- Accuracy on your task, using the score your team already trusts: task accuracy, defect recall or word error rate.
- Consistency at volume: the format, terms or labels hold across thousands of calls, not only in a demo.
- Cost and speed, measured on the hardware you'll run it on, reported as cost per call and P50/P95 latency.
- Control over where it runs: your cloud, a private network, on-premises hardware or a device.
- No restart when a better base model ships: you keep the evaluation suite and pipeline, so you can score a new model and re-train only if it's worth it.
We don't quote accuracy figures before your baseline exists.
What you receive
| Deliverable | What's in it |
|---|---|
| 1. Fit memo | Fine-tune or not (and the cheaper option if not), candidate base models and licences, data readiness, cost drivers, risks |
| 2. Evaluation suite and baseline | A test set your expert approved, the metrics for your model type, regression checks, your current model's scores |
| 3. Training dataset and pipeline | Versioned examples (text, labelled images or transcribed audio), personal data handled, a record of where each came from, and the rebuild script |
| 4. Trained model | Weights or adapter, configurations and a log of every run |
| 5. Evaluation report | New vs current model: task score, regressions, latency, cost per call, known failure cases |
| 6. Deployment package and handover | Model files for your runtime, a reference serving configuration, a model card, a retraining runbook and a handover session |
A Fit Sprint delivers 1 and 2. A Build delivers all six.
You own the trained model, the dataset and the evaluation set once the work is paid for; we keep our own reusable tooling, and the base model's licence still applies. Need it connected to your product? We can scope that integration into the Build.
What we need from you
| From you | Language | Vision | Speech |
|---|---|---|---|
| A defined task | “Route tickets into 14 queues” | “Flag four defect types on panel images” | “Transcribe support calls, including product names” |
| Examples, and the right to train on them | Logs or labelled examples | Labelled images, or images plus your experts' time to label | Audio with accurate transcripts, with consent |
| Your current setup | Model, prompt, sample outputs | Current model or inspection process, and its miss rate | Current speech service and sample transcripts |
| A domain expert, available every week | Approves labels and the bar | Confirms what counts as a defect | Checks transcripts of key terms |
| Where it must run | Cloud, private network, on-premises | Line camera, edge device or server | Cloud, on-premises or device |
For all three: a cloud account for training compute, and a named owner for the model after handover.
Not included
Hosting and running the model after handover (your team runs it, with the runbook we hand over) · labelling at scale beyond the agreed plan · building the RAG system (RAG Development) · building a model from scratch for a prediction problem (Machine Learning Engineering) · buying third-party licences such as YOLO Enterprise (we tell you if you need one) · promising accuracy before your baseline exists.
Which models, and where they run
Open-weight models in your environment. Most work uses open-weight models trained in your own cloud account, a private network or on-premises, so text, images and audio stay with you. We check each candidate's licence in the fit memo.
Managed fine-tuning, where it fits (checked November 2026):
- Language: Google supports tuning for Gemini 3.5 Flash, 3.5 Flash-Lite, 3.1 Flash-Lite and the 2.5 family. AWS Bedrock lists Amazon Nova, Meta Llama 3.1–3.3 and Claude 3 Haiku, its only Anthropic model. OpenAI stops new fine-tuning jobs on 6 January 2027.
- Vision: Microsoft Foundry tunes selected GPT models with images, skipping images of faces or people.
- Speech: Microsoft custom speech and Google speech adaptation, usually tried before fine-tuning.
A model tuned on a managed platform runs on that platform, and its weights stay with the provider. The fit memo compares both routes.
Data handling: personal data is removed or masked before training. We sign a mutual NDA before you share data, and your data is used only to build your model.
How an engagement runs
Fit (go/no-go; Gate 1) → Baseline (evaluation suite you approve) → Data (curate, label, hold out a test set) → Train (adapters first, two or three base models) → Prove (held-out set, regression checks, real production samples, cost and latency; Gate 2) → Hand over (package, document, and your team re-runs the pipeline once). Weekly written report and decision log throughout (How We Work).
How you buy it: free fit call → Fit Sprint → Build. The Fit Sprint and the Build are each priced per scope. The full stage-by-stage breakdown is on the custom AI model development page.
What can go wrong, and how we catch it
| Risk | How it's caught |
|---|---|
| Better at your task, worse at something it used to do well | Regression checks scored at Gate 2, before you accept anything |
| The training data is weaker than expected | Data checks, and a held-out test set that training never sees |
| A vision model works on clean test images and fails on the line | The Prove stage scores real production images, including glare, blur and rare cases |
| A speech model learns your terms but gets worse on everyday words | Word error rate reported overall and on key terms, side by side |
| A better base model ships next quarter | You keep the evaluation suite and pipeline: score it, and re-train only if it's worth it |
How to judge us
The method, the gates and the deliverables are in the contract.
| Question | Crescent AI | Self-serve fine-tuning platform | Generalist AI agency |
|---|---|---|---|
| Decides whether to fine-tune | Yes, in a written go/no-go | You decide | Varies |
| Language, vision and speech models | Yes, one process | Usually language only | Varies |
| Managed customization tested first | Yes | No | Rarely |
| Licence review (incl. YOLO, Llama, Gemma) | Yes, at the Fit stage | No | Varies |
| Evaluation built before training | Yes, a bar you approve | You build it | Often mentioned |
| Best when | You need the decision, the proof and the handover | You have clean data, a metric and an engineer to run training | You need a broad AI application built |
Who leads this work
Yash Amin, founder of Crescent AI. About Crescent AI
Common questions
How much does it cost to fine-tune a model?
It's priced per scope. The Fit Sprint and the Build are each scoped to your task: data readiness, model type and size, and how much evaluation you need. Training compute runs on your own account.
How much data do we need?
For a narrow language task, a few hundred good examples is a practical start. For vision, it depends on how many defect types you have and how different they look, which we work out from your images in the Fit Sprint. For speech, enough transcribed audio to cover your key terms and accents.
Which LLM is best for fine-tuning?
It depends on the task, the licence and where the model must run, so we compare two or three candidates on your test set.
Is fine-tuning better than RAG?
They fix different problems. RAG gives a model your facts; fine-tuning changes its behaviour. Many systems use both.
Can ChatGPT, Claude or Gemini be fine-tuned?
OpenAI stops new fine-tuning jobs on 6 January 2027. On AWS Bedrock, the only Claude model listed for fine-tuning is Claude 3 Haiku. Google supports tuning for several Gemini 2.5, 3.1 and 3.5 Flash models. (Checked November 2026.)
Can we use YOLO commercially after fine-tuning it?
Only with an Ultralytics Enterprise Licence, unless your whole project is open-sourced under AGPL-3.0. We'll tell you if that applies and compare alternatives.
Is fine-tuning Whisper better than custom vocabulary or hotwords?
Try custom vocabulary first; it's cheaper. Fine-tuning helps when the problem is accents, audio conditions, or a long list of ambiguous terms.
How long does fine-tuning take?
It depends on your data and how much evaluation you need, which the Fit Sprint sizes for your task. Training runs are often the short part; most of the work goes into data and evaluation.
Do we own the fine-tuned model?
Yes. You own the trained weights or adapter, the dataset and the evaluation suite once the work is paid for. The base model's licence still applies, and the fit memo lists its terms.
Do you host the model afterwards?
No. We hand it over with a runbook, and your team runs it.
Bring us the examples your model keeps getting wrong.
Text, images or audio. We'll tell you whether fine-tuning is the fix before you spend on it, and if it is, prove it on your own test set.
Mutual NDA before you share data · Fixed-scope Fit Sprint · No training spend before your go/no-go
Part of Custom AI Model Development · Model Distillation · Embedding Model Fine-Tuning · Fine-Tuned Model Migration