Rixven Inc. — founded 2026
The Next Era
Rixven is a two-person research company building an original model architecture from first principles, and shipping the specialist tools that come out of that work along the way.
loading=eager
fetchpriority=high
/>
Both are true at once, and neither subsidises the other.
Years — the research
A transformer-replacement architecture built around a resonance field instead of an attention matrix — no KV cache, no quadratic lookup, state that persists by phase rather than by storage. An eleven-phase roadmap. We are early in it and publishing as we go.
The full architecture →Quarters — the product
Developers now spend more hours reviewing AI-written code than writing their own, and the failures that survive review are confident, fluent, and wrong. Kerven Vega 1 is a small, open-weight model whose only job is catching them.
What it does →Sponor will not produce a shippable model soon. That is not a hedge; it is the honest timeline for architecture research done properly, and we would rather say so than imply otherwise.
Kerven Vega 1 exists on its own terms, not as a stopgap. It is a real, validated gap: nobody has built a small model trained specifically to catch the way machine-written code fails, which is different from how human code fails. It funds the company. It also keeps us building for people who use the thing today, while the longer work continues underneath.
Kerven Vega 1's training pipeline is JAX and Flax on a TPU v5e-8, granted through Google's TRC programme.
The KV cache is not a bug in the transformer. It is the transformer, presenting its bill.
If a model's context is a resonance field, its maximum useful length is not a hyperparameter. It falls out of the spectral gap.
The age is not a marketing device. It is who wrote the code, and we think that is worth saying plainly rather than hiding or overplaying.
More about the company →