First, the terms
- Hosted (closed) models: you use them through an app or API run by the provider, e.g. OpenAI (ChatGPT), Google (Gemini), Anthropic (Claude). You can't see or download the model itself.
- Open-weight models: the trained model is published, so you can run it on your own servers or a cloud provider of your choice. Well-known families include Llama, Mistral, Qwen, DeepSeek, Gemma and Kimi. Licences vary, so check commercial terms.
Side-by-side
| Hosted (ChatGPT-style) | Open-weight, self-hosted | |
|---|---|---|
| Getting started | Minutes: sign up, use the API | Days to weeks: set up servers or a host |
| Quality on hard tasks | Usually the best available | Close behind, varies by task |
| Cost at low volume | Low: pay per use | Higher: servers cost even when idle |
| Cost at high volume | Grows with every request | Can be much cheaper per request |
| Data control | Data goes to the provider (check their terms) | Data can stay on infrastructure you control |
| Customisation | Prompts, some fine-tuning | Full control, including fine-tuning |
| Maintenance | Provider handles it | You handle updates, scaling, security |
When a hosted model is the better choice
- You're starting out and want results this week.
- Your volume is low or unpredictable.
- You need the very best quality for complex reasoning, writing or coding.
- You don't have someone to run and secure AI infrastructure.
When an open model makes sense
- Sensitive data that you can't or don't want to send to a third party (health, legal, financial, government).
- High, steady volume: thousands of documents or conversations a day, where per-request fees add up.
- A narrow, repeated task (classifying tickets, extracting fields from invoices) where a smaller tuned model does the job well.
- Offline or on-premise needs, such as a factory, hospital or remote site.
Most businesses don't have to choose forever. A common pattern is to start with a hosted model, learn what works, then move high-volume or sensitive tasks to an open model.
The hidden costs of self-hosting
- GPUs or GPU cloud time, which is the biggest cost and is paid whether or not anyone is using the model.
- Engineering time for setup, monitoring, scaling and updates when better models come out.
- Security: a model server is another system that needs patching and access control.
- Evaluation: you need your own tests to know whether the open model is good enough for your task.
Build so you can switch
The model landscape changes every few months. Whatever you choose, keep the model behind a single layer in your software so you can swap providers, or move from hosted to open, without rewriting your application. Keep a small set of real test questions with known good answers so you can compare models quickly when you consider a switch.
A simple decision path
- 1. Prototype the task with a hosted model.
- 2. Measure quality, monthly cost and how sensitive the data is.
- 3. If cost or data control becomes a problem, test one or two open models on the same task.
- 4. Move only the tasks where the open model is good enough.
Open models are now a serious option, not a hobby project. For most small businesses, hosted models are still the fastest way to start. For high volume or sensitive data, open models can cut costs and keep data in your hands. We help businesses choose, test and integrate AI models into their own software, hosted or self-hosted, designed so you can change models later.
Talk to us about AI automation or custom software, or compare the main hosted assistants in ChatGPT vs Gemini vs Claude.
