Comparison
Fine-tuning vs training your own model
The difference between fine-tuning, post-training, and training from scratch, and which one you actually need.
By James Drayson
In short
Fine-tuning lightly adapts an existing model to a task; post-training deeply specialises a base model on your domain while preserving its general ability; training from scratch builds a model from zero. For most enterprises, post-training, what Locai does with Forget-Me-Not, gives domain expertise and ownership without the cost of training from scratch.
The three approaches, defined
- Fine-tuning: A light adaptation of an existing model to a narrow task or style; quick, but shallow and prone to forgetting.
- Post-training: Deeper specialisation of a strong base model on your domain, language, and workflows, preserving general reasoning.
- Training from scratch: Building a model from zero, maximum control, but very high cost and time.
What Locai does, and why
Naive fine-tuning causes catastrophic forgetting: the model gains your task but loses general ability. Locai's Forget-Me-Not framework post-trains a strong base on your data while preserving its general capabilities, producing a domain expert that still reasons well, and can keep learning continually. You own the resulting weights.
Fine-tuning vs post-training vs from scratch
| Post-training (Locai) | Fine-tuning | From scratch | |
|---|---|---|---|
| Depth of specialisation | Deep | Shallow | Deep |
| Preserves general ability | Yes (Forget-Me-Not) | Often degrades | Yes |
| Cost | Moderate | Low | Very high |
| Continual learning | Yes | Limited | Yes |
| You own the result | Yes | Often not (hosted) | Yes |
What this looks like with Locai
Where the comparison points toward running AI on your own hardware, this is what that looks like delivered as a product.
Locai Labs builds Locai One, an on-prem AI appliance. It is one machine that arrives with everything already in it: the hardware, our open-weight Locai Juno models, and Locai OS, the operating system that serves the models and handles users, access and monitoring. You plug it into a mains socket and your network, and your team is working in about 15 minutes. No cloud account, no per-token bill, and nothing leaving the building.
The reason a data-centre-class model fits in a box on your floor is SPACE, our compression algorithm. Instead of asking how much of a model can be cut while keeping it broadly similar, SPACE asks what the model needs to be good at, preserves the subnetworks behind those capabilities and strips back the rest. The result is a smaller specialist rather than a shrunken generalist, tuned to the exact hardware it ships on.
Locai One starts at £29,950 for a team, and Locai One Pro at £49,950 for an organisation, bought once and owned outright. Both run air-cooled on standard mains power and work fully air-gapped. Any compatible open-weight model runs alongside Juno, and if you need a model trained on your own proprietary data we can post-train one and deploy it on the same machine.
Frequently asked questions
What's the difference between fine-tuning and training?
Fine-tuning lightly adapts an existing model; training (from scratch) builds one from zero. Post-training sits between, deeply specialising a base model while keeping its general ability.
Is fine-tuning enough?
For narrow tasks, sometimes. For deep domain expertise without losing general reasoning, post-training is stronger, and it avoids the catastrophic forgetting that fine-tuning can cause.
What is post-training?
Specialising a strong base model on your domain, language, and workflows. Locai uses Forget-Me-Not to do this while preserving general capability.
Do I own a fine-tuned model?
Often not, fine-tuning on a hosted API usually keeps the result on the vendor's platform. With Locai's post-training you own the weights outright.
Book a sovereign AI briefing
A 30-minute session on owning your model: deployment options, the data path, and a clear cost range for your use case.