Small language models: when smaller AI is the smarter choice
What small language models are, where they match bigger models, where they don't, and how businesses use them for speed, cost and privacy.
How open-weight and proprietary AI models compare on capability, cost, control, privacy, licensing and operations, and how to choose for your project.
Proprietary AI models, accessed through providers' APIs, often lead on capability, need no infrastructure and come with support. Open-weight models can be downloaded, run on your own infrastructure, fine-tuned freely and kept fully under your control, but need hosting and expertise. Licences vary, so read them. Many businesses use proprietary APIs for complex tasks and open models for privacy-sensitive or high-volume workloads.
When building AI features, one key decision is whether to use a proprietary model through an API or run an open-weight model yourself. Both are viable, and the best answer often combines them.
Proprietary models are developed by companies and accessed through their APIs or products. You don't get the model weights. Examples include leading models from OpenAI, Anthropic and Google.
Open-weight models have their trained weights published, so anyone can download and run them, subject to the licence. Examples include model families from Meta, Mistral and others. "Open source" is often used loosely; some models release weights but not training data or code, and licences vary.
| Proprietary (API) | Open-weight (self-hosted) | |
|---|---|---|
| Capability | Often at the frontier | Strong and improving; varies by model |
| Setup | Fast | Needs infrastructure and expertise |
| Cost model | Pay per use | Infrastructure and staff costs |
| Data control | Depends on provider terms | Full control in your environment |
| Customisation | Limited fine-tuning options | Full fine-tuning |
| Updates | Provider manages, sometimes changing behaviour | You choose when to update |
| Support | Provider support plans | Community or vendors |
Open-weight models come with different licences. Some are permissive; others restrict commercial use above certain scales, require attribution or prohibit certain uses. Read the licence before building a product on any model, and keep records of which version and licence you used.
Many providers offer business terms where your data isn't used for training, with retention controls and regional data options. Review the terms, data processing agreements and security certifications. For highly sensitive data, self-hosting may still be preferred.
Managed platforms that host open models for you reduce this burden while keeping some flexibility.
At low or variable volumes, APIs are usually cheaper. At high, steady volumes, self-hosting can be cheaper, but include engineering time. Our guide to AI cost optimisation covers ways to reduce costs either way.
Many teams:
Test candidate models on your own evaluation set before deciding. Understanding the different branches of AI helps too; see our guide to machine learning, deep learning and generative AI.
Benchmarks published by model developers are a starting point, but your own data matters more. Build a test set of real tasks, define what a good answer looks like and compare models on accuracy, speed and cost. Repeat the comparison when new versions appear, because the gap between models changes quickly.
Design your application so the model sits behind a clear interface. Keep prompts, evaluation sets and business logic separate from provider-specific code. This lets you switch models, or use several, as capabilities and prices change.
Model capabilities, prices and licences change quickly. Schedule a review every few months to check whether a different model now offers better value.
Our AI solutions team evaluates proprietary and open models against your data and requirements. Our web app developers integrate APIs or deploy self-hosted models, and our app development service builds flexible architectures that let you switch models.
The weights may be free to download, but hosting costs money and licences may restrict some uses.
Some are very capable, especially for focused tasks. Frontier proprietary models often lead on the most complex tasks.
Many providers offer business terms that protect data. Review terms carefully, or self-host for maximum control.
This guide is general information, not legal, tax or financial advice, and details such as prices, features and rules can change. Product and company names mentioned are trademarks of their respective owners and are used only to identify those products; UK Technology Hub is not affiliated with or endorsed by them.
Tell us about your project and get a free, fixed quote.
Get a free quoteWhat small language models are, where they match bigger models, where they don't, and how businesses use them for speed, cost and privacy.
How the main branches of AI relate to each other, what each is good at, and which approach fits common business problems.
The real differences between native, hybrid and cross-platform mobile apps in performance, cost, user experience and long-term maintenance.