Research
The thinking behind what we build.
Our product pages tell you what Run BiOS does. These tell you why it is built the way it is, what we think the hard problems are, and where we are not finished. Every external figure carries a live link to its source, and anything that is our own judgement is written as judgement.
The BiOS Conscious Loop: giving an enterprise model a memory
Enterprise AI spending is rising steeply while the systems being bought get no better at the buyer’s work. We argue the constraint is structural rather than a matter of model quality, describe the six-stage loop that fixes it, and state the failure modes it has to survive.
The BiOS training engine: single-device model code, declared parallelism
The guarantees an enterprise asks for cannot be made by a layer that does not control what breaks. How the engine is organised, the five ways a run can be split, and the discipline we would defend hardest: refusing to degrade quietly.
The BiOS inference engine: serving as a cache-management problem
An agent revisits the same material constantly, which makes repetition the largest opportunity in the workload. The four cache tiers that exploit it, why exactness is what makes caching safe to leave on, and why one image across accelerator generations is a requirement on-premise.
Adaptive inference: matching the model to the work
Work is not uniformly hard, but capability is usually priced as though it were. Why the collapse in unit price promoted allocation rather than solving it, the measurement problem underneath, and why allocation and specialisation converge.
Sourced. Every third-party figure is attributed to a named publication with its date and a working link, listed at the foot of the page it appears on.
No borrowed numbers. We do not publish a performance figure without a method beside it. A headline multiplier with no harness attached is advertising.
Honest about gaps. Each paper names what it has not solved. A research page that lists only strengths is a brochure with footnotes.