Our products
Both products start from the same idea: AI should not depend on one machine or one place.
Dune
Distributed LLM inference for critical AI infrastructure
Dune pools machines an organization already owns into one cluster that serves language models through an OpenAI-compatible API.
It spreads the work across ordinary machines, with CPUs or GPUs, and keeps running when some of them are lost.
What Dune does
- Any machine can take any request and pass it to a copy of the model that is running.
- No machine is in charge. The machines find each other on their own, and if one of them or a whole site is lost, the rest keep serving without anyone stepping in.
- A request sent to a lost or slow copy is tried again on another, so people get an answer, not an error.
- A new model version reaches one copy first, and Dune goes back to the previous version on its own if the new one does worse.
- Optional federated learning: each site can improve a model with its own data, and that data never leaves the site.
Models
Dune serves open language models, including versions an organization has fine-tuned on its own data. Several adapters can share one base model, and a model too big for one machine can be split across several.
Built on proven engines
Dune does not rebuild inference from scratch. It runs established open-source inference engines and adds a layer above them for resilience, security and day-to-day operations.
What it serves today
Today Dune serves chat-style language models only. It does not serve embedding, vision or speech models.
Why it matters
Critical AI infrastructure has to be robust, resilient and always available, even when part of it fails. Dune is built for that.