Local-first does not mean local-only

This site doesn't talk about sovereign AI in the abstract — it runs on it. It's worth showing exactly how, because "local-first" is often misread as "never use the cloud," and that's not the stance.

The real hierarchy

Production runs today with three models in priority order: a 27B primary model on the main GPU, an 8B secondary always active on the second GPU, and a small fallback model on CPU for when no GPU is available. Each tier has a clear role — the primary handles most reasoning tasks, the secondary covers load spikes or simpler tasks, and the CPU one is the guarantee that the system never goes down completely.

The cloud exists in this architecture, but as explicit fallback, not as a silent default. That matters: a system that "sometimes" sends your data to an external API without you knowing is not sovereign, even if it has a local model installed. The difference is not having your own hardware — it's that the default path is local and the exception is documented and visible.

Why the difference matters

Running models locally is not a cost-free ideological stance — it has concrete, measurable advantages: near-zero marginal cost once the hardware is amortized, no rate limits imposed by an external provider, and full control over when and how the model you are using gets updated. It also has a real cost: infrastructure maintenance, setup time, and a capacity ceiling compared to commercial frontier models.

The honest conclusion

Local-first means that by default the data stays on your infrastructure, and that when it leaves — because sometimes it makes sense for it to leave — you decided it would leave, not a default nobody looked at. That is the bar this site uses to judge any tool it reviews, starting with its own.