Better architecture beats bigger clusters.
Valerois was founded on one conviction: models do not need frontier-scale compute if their core design is right. We build Nova on top of that belief and research the models that will power it next.
Heimdall
A hybrid sequence architecture with a small cache, 1-bit feed-forward layers and a method to convert existing models. It runs and trains on one consumer GPU.
Nova
The AI workspace that plans, writes, runs and checks your work, on your desktop or in your browser. This is what most people come to Valerois for.
Efficiency is a research direction.
Constant or near-constant memory opens doors that a growing cache keeps shut: long-running agents, private on-device assistants and affordable research. Nova is where that shows up for people today. These are the rules we work by.
Honest numbers
We publish the table where we are behind, and mark plans as plans.
One GPU first
If an idea only works on a cluster, we are not ready to claim it.
Ship it
Research that never reaches people is half a result. Nova is how it reaches them.
Private by default
No trackers on our site. Local runtimes and your own keys on Nova.
One line. Down, up, down.
Valerois joins valeur (worth, valour) and roi (king). The mark is one unbroken stroke that goes down, up and down again: a V that turns into a Λ. It is the first letter of the name, and it is also the shape of a signal that keeps its own course. Nova, the star, is the product: a point of light small enough to sit next to a text box.
What is done, and what is not.
Nova
Desktop app (Windows, macOS beta, Linux) and web app. Plans from free to Max, Supernova mode, skills, connectors, local models.
Heimdall-Bit 98M · 337M
Hybrid backbone, latent-free 1-bit FFN, BitFlip optimizer, Triton kernels and an OpenAI-compatible server. 337M parameters trained on 3.7B tokens on one consumer GPU.
Scale study
Binary-FFN hybrid and a dense baseline at 300M to 1B parameters on at least 10B tokens, several seeds. Depends on compute.
Conversion of larger models
Convert Qwen3 1.7B to 8B with a 4-bit teacher; measure long-context quality and serving cost.
Go on. Try it.
Nova is free to start. No card, no setup afternoon. Or talk to us about the research.