ChatGPT alternatives
By Daniel Reeve · tested and updated September 2026
Open-source AI can be attractive when you want more control, local deployment, lower marginal cost or customization. But it is not automatically easier or cheaper than hosted AI.
Quick answer
Do not read “open” as free-and-ready. Read the licence column: three different licence tiers appear in this family, and only one of them permits commercial use.
Decision map
Developers, privacy-sensitive teams, regulated environments and companies with technical ownership.
Beginners who want a polished app with minimal setup.
Run your real task through a hosted model and a local model, then compare quality, speed, cost and maintenance burden.
Licence by licence
| Area | Useful for | Watch out for |
|---|---|---|
| Local privacy | Data can stay on your machine or infrastructure if configured correctly. | Plugins, logs and connected tools can still leak information. |
| Cost control | Self-hosting can reduce variable API costs at scale. | GPU, engineering and monitoring costs are real. |
| Customization | Teams can fine-tune, prompt-wrap or deploy specialized models. | Custom models require evaluation and guardrails. |
| Offline or restricted environments | Local AI can work where cloud tools are not allowed. | Quality may be lower than top hosted models for some tasks. |
| Developer learning | Great for understanding models, retrieval and inference. | Not the fastest route for normal users who just want answers. |
Verified open weights
“Open source” is used loosely in AI writing. The licence column is the one that matters, and the three tiers below are not interchangeable. Read from Mistral model weights documentation on 2026-09-10.
| Model | Parameters | Licence |
|---|---|---|
| Mistral 7B | 7.3B | Apache 2.0 |
| Mixtral 8x7B | 46.7B total / 12.9B active | Apache 2.0 |
| Mixtral 8x22B | 140.6B total / 39.1B active | Apache 2.0 |
| Mistral Nemo | 12B | Apache 2.0 |
| Ministral 8B | 8B | Mistral Research License |
| Mistral Large | 123B | Mistral Research License |
| Pixtral Large | 124B | Mistral Research License |
| Codestral | 22.2B | Mistral Non-Production License |
Before you self-host
First: does the licence allow your use? Apache 2.0 does, a Research License does not, and that answer costs nothing to establish and everything to get wrong.
Second: is control worth the running cost? At low volume, self-hosting is usually more expensive than an API, not less — and the safety tuning becomes your responsibility, as Mistral's own documentation states.
Sources
Every figure on this page was read from the official documentation below on 2026-09-10. Prices, limits and model names change without notice — the source is authoritative, this page is not.
FAQ
Only the Apache 2.0 ones, strictly speaking: Mistral 7B (7.3B), Mixtral 8x7B (46.7B total / 12.9B active), Mixtral 8x22B (140.6B / 39.1B active) and Mistral Nemo (12B). Ministral 8B, Mistral Large (123B) and Pixtral Large (124B) ship under the Mistral Research License, and Codestral (22.2B) under a Non-Production License — those are downloadable, not open source (Mistral docs, checked 2026-09-10).
The Apache 2.0 models, yes. The Research License and Non-Production License models, no — that is precisely what those licences restrict. This is the distinction most “best open source AI” articles skip, and it is the one that creates legal exposure.
Roughly, parameter count drives memory. A 7B model runs on a good consumer GPU; a 123B model does not. The active-parameter figures for the Mixtral models matter here: 8x22B has 140.6B total parameters but 39.1B active, which changes the compute story without changing the memory you need to hold it.
For many everyday tasks the gap is smaller than the discourse suggests; at the frontier it is real. The honest framing is that you are not buying equivalent quality, you are buying control — over data, over hosting, over whether the model changes underneath you next week.
Two things. The safety tuning is yours to build, as Mistral's own documentation states plainly. And the running cost is yours too: at low volume, self-hosting is usually more expensive than an API, not less. Self-host for control and data residency, not to save money.