Open models are the ones you can download and run yourself. They trail the closed frontier, by much less than they used to, and for a meaningful set of jobs the trade is worth it.
"Open weights", not "open source"
Worth getting right, because the distinction has practical consequences.
Nearly every model called open-source releases weights — the trained parameters. You can run them, fine-tune them and deploy them. What is generally not released is the training data and the full pipeline, which is what "open source" means in software.
So the accurate term is open weights. Licences vary too: some are genuinely permissive, some restrict commercial use above a size threshold. Read the licence of the specific model rather than assuming.
The main families
| Model | From | Notable for |
|---|---|---|
| DeepSeek | DeepSeek | Reasoning and code, very low cost |
| Llama | Meta | The broadest ecosystem and tooling |
| Qwen | Alibaba | Strong multilingual performance |
| Mistral | Mistral AI | Efficient models that punch above their size |
| Gemma | Small models designed to run locally |
The gap to the closed frontier keeps narrowing. On reasoning and code specifically, the best open models are close enough that the difference does not show on many real tasks.
Why you would actually want one
Data that cannot leave. The strongest argument by a distance. Healthcare, legal, defence, anything under strict data residency rules. Self-hosting removes the question entirely rather than answering it contractually.
Cost at volume. Once you are processing serious quantities, running your own model on your own hardware becomes cheaper than per-token pricing. The crossover point is higher than enthusiasts claim, but it is real.
No vendor risk. Nobody can deprecate your model, change its behaviour overnight, raise the price, or refuse you service. If it works today it works in three years.
Fine-tuning. Genuinely adapting a model to your domain and your data, rather than stuffing context into a prompt.
Offline. Air-gapped environments, edge devices, anywhere without reliable connectivity.
Why you probably still want a closed model too
The frontier is still closed. For the hardest reasoning, the best writing and the most reliable instruction-following, the closed models remain ahead.
Total cost is not just the API bill. Self-hosting means GPUs, ops, monitoring and someone whose job includes keeping it running. Teams routinely underestimate this and end up spending more than the API they were avoiding.
The tooling around closed models is better. Integrations, documentation and solved problems.
Multimodality lags. Open models handle text and increasingly images; the breadth of closed offerings across voice, video and documents is wider.
Running one locally
More approachable than most people expect for the smaller models.
- Small models (roughly 7–8B parameters) run on a decent laptop, especially with Apple silicon or a modern GPU. Genuinely useful for summarising, extraction and simple drafting.
- Mid-size (30–70B) needs a serious GPU or heavy quantisation.
- The largest open models need infrastructure, not a workstation.
Quantisation — compressing weights to lower precision — is what makes local running practical. It costs some quality for a large reduction in memory. Most local deployments use it.
Realistic expectation: a small local model is a capable assistant for well-defined tasks, not a replacement for a frontier model on hard work.
Choosing
- Do you have a hard data-residency requirement? If yes, open weights, and the decision is made.
- Are you processing enough volume that per-token pricing hurts? If yes, model the total cost including operations before switching.
- Otherwise, start closed and keep an open model available for the specific jobs where cost or privacy makes it the better tool.
Most organisations end up with both, which is the sensible outcome rather than a failure to decide.
Common questions
What are the best open source AI models? DeepSeek for reasoning and code, Llama for ecosystem breadth, Qwen for multilingual work, Mistral for efficiency, Gemma for running locally.
Are open models as good as ChatGPT? Close on reasoning and code. Behind on the hardest tasks, on prose, and on the polish of the surrounding product.
Are they really free? The weights are. Running them is not — you pay in hardware and operations instead of per token.
Can I run one on my laptop? Small models, yes, with a modern machine. Larger ones need serious hardware.
Is open source AI safe for business data? Self-hosted, yes — the data never leaves your infrastructure, which is the main reason regulated industries choose them.
What does quantisation mean? Compressing the model to lower numerical precision so it fits in less memory. Some quality cost, large practical gain.
Can I fine-tune an open model? Yes, and it is one of the main reasons to choose one. Not possible in the same way with closed models.