The rest of this section is about renting AI. This page covers the other option: models you download and run yourself, on hardware you control. Five years ago that was a research hobby. Today the open-weight world is serious enough that for some workloads it’s the sensible default, and for some regulated work it’s the only answer a compliance team is comfortable with.
The families that matter
Five well-backed families cover most of the open-weight field, each with a range of models from small to large:
- Llama (Meta): the most widely deployed family, with the broadest tooling support. Custom licence, permissive for normal businesses but with conditions worth one careful read. There’s more in the Meta guide.
- Mistral (Mistral AI, France): Europe’s leading lab, popular where an EU headquarters and data handling matter, with permissive licensing on many models and strong small models. More in the Mistral guide.
- DeepSeek (China): the family that shocked the market in early 2025 by matching far more expensive models at a fraction of the training cost, published under permissive terms. Capability per pound is the draw. For some firms, its origin and their risk committee’s view of it will be part of the assessment.
- Qwen (Alibaba): a broad, fast-improving family with strong multilingual performance, widely used and mostly permissively licensed.
- Gemma (Google): Google’s open family, smaller than its Gemini flagships, well documented and convenient for teams already using Google’s tools.
The shortlist at a glance:
| Family | Backer | Licence posture | Known for |
|---|---|---|---|
| Llama | Meta | Custom community licence | Widest deployment, deepest tooling |
| Mistral | Mistral AI (France) | Largely Apache 2.0 | Small-model quality, EU comfort |
| DeepSeek | DeepSeek (China) | MIT-style | Capability per pound |
| Qwen | Alibaba | Largely Apache 2.0 | Multilingual breadth |
| Gemma | Open weights, Google terms | Documentation, ecosystem fit |
Names and version numbers change constantly, but the families and their characters stay fairly stable. Anything this page said about “the current best” would be out of date within a quarter, which is a lesson in itself about how firms should decide.
Three reasons to run your own
There are three good reasons to run a model yourself, and one big cost.
Data control. Prompts, documents and outputs never leave infrastructure you own. For workloads where sending client data to any third party is unacceptable, or just awkward and expensive to paper over, this settles it. It’s why the Software Factory offers it as an option.
Cost at volume. APIs charge per use, and your own hardware costs the same every month whether you use it or not. High, steady volume eventually reaches the point where owning is cheaper. The sum is dull but simple: expected tokens, hardware rental, and an allowance for the person who keeps it running. Do it before believing either answer.
Independence. No provider can reprice, rate-limit, retire or change the model underneath you. For systems meant to run for years that stability is worth something, and it adds to the case for being able to switch providers made across this whole section.
Against all three sits the cost nobody advertises: you become the host. Updates, security, monitoring, capacity and backups are the same discipline any production system needs. That’s fine if you have it, and a liability if you don’t.
The middle path: hosted open models
Most firms curious about open weights should start with hosted open models, which sit between “pay OpenAI per token” and “run your own servers”. The major clouds and specialist providers serve Llama, Mistral, DeepSeek and Qwen per use, sometimes very cheaply. You get the choice of models and some of the independence without becoming a hosting operation. It’s also the cheap way to check whether an open model can handle your work before any conversation about hardware.
The checklist
Before you commit to running your own model, check that you have:
- Read the licence and cleared it for your intended use
- Tested the model on your own tasks against a hosted version first
- Calculated the volume economics, with the crossover point written down
- Named who owns the hosting, with monitoring, updates and backups in the plan, not the appendix
- Assessed data protection as you would for any processor: self-hosting moves the obligations, it doesn’t remove them (governance guides, DPIA starter)