Tuna was founded on a simple obsession: speed.

The benefits of speed need no explanation. The only question is how.

We start with our quality-first commitment. Speed means absolutely nothing if the output is garbage. Some inference optimizations have no effect on quality, while others make the model dumber. So we rigorously evaluate optimization techniques to ensure they preserve the model's original quality and capabilities. For example, we do not quantize models below their original precision. We are not willing to sacrifice even an ounce of intelligence.

Within these constraints, we engineer inference to run as fast and efficiently as possible. We apply good old elbow grease and traditional engineering techniques. We pair this with Bluefin (our self-improving autoresearch system) to optimize every single aspect of the stack and enable rapid deployment on diverse hardware.

We are building heterogeneous compute infrastructure to maximize speed per dollar. Traditional inference systems use GPUs as generalists across multiple stages, but this creates a mismatch because each stage places different demands on compute and memory. By matching each stage to specialist hardware, we can provide access to faster AI at a reasonable price.

We believe the freedom to access AI is part of the American Dream, and that freedom should be shared with everyone around the world. US regional routing is the only option, at no extra cost. People should be able to use powerful AI without giving up their privacy. Zero data retention (ZDR) is the default for everyone. We are building Tuna to turn that dream into opportunity for all of humanity.