AI research & development · Texas
Models and hardware, made efficient together.
Ember Compute is an AI R&D company in Texas. We work on the line between a model and the machine it runs on: how much compute a model really needs, how much of the hardware it actually uses, and how to close that gap on devices people own.
Model-hardware compute efficiency
Efficiency is a property of the pair, not of either part. We work on all three places where it is won or lost.
Compute-efficient models
Small, domain-tracking models that reach parity on the current task and adapt as the task drifts, instead of carrying every tail of knowledge in resident weights.
Memory bandwidth follows active parameters. Smallness is the enabler, not a compromise.
Runtimes that learn while they serve
Two-process loops: one face serves at frame rate, the other trains on the same stream, and weights are hot-swapped mid-stream without a dropped frame.
Latency is measured at p99, on the frame thread, with a frozen control beside it.
Hardware shaped by the model
Mixture-of-experts models made capacity the lever: the whole model must be resident, but only a few experts move per token. When no one machine has the memory, every machine on the network becomes one pool, and only activations cross the network.
Fully local to the building. Nothing leaves it.
Two projects, one thesis
Janus learns while it serves on one device. WDMA serves a model bigger than any one device. Each has its own page.
Janus
A video world model that learns and infers at the same time, on the same machine, against the same live stream. The served output is surprise: an unlabeled, on-device "I don't know" signal.
WDMA
Wireless Direct Memory Access: serve a mixture-of-experts model bigger than any one machine across the memory of the machines you already have, on the network you already have, from a web page. Weights stay where they are. Activations move.
Ember Compute LLC
An AI research and development company based in Texas, founded in 2025. Our focus is model-hardware compute efficiency: closing the gap between what a model needs and what the machine underneath it delivers.
We measure against baselines a demo cannot fake
Every number we publish is reported against a frozen control, an ablation, or persistence. Raw error on its own is not a result.
We publish the nulls
A change that does not move the metric is still a finding, and it is written up with the same care as a win.
We build for devices people own
The target is the machine on the desk, in the bag, or on the robot. Not a datacenter, and not a subscription.
Talk to us
Engineers, partners, and prospective customers of either project are welcome.
Texas, USA
Ember Compute LLC