One thing you can use today.
Most of what SHV is building is not ready, and we would rather point you at the part that is than describe endpoints that do not exist yet.
Available now: the long-horizon layer.
Eunoia-4b-mini is not a model. It sits between your prompt and whatever model you already run, and keeps intent, constraints and reasoning structure intact as generations get long — the point where most agents quietly stop following the instruction.
It is open source, it works on any base, and the reference build is published so you can judge the claim before wiring anything up.
Wrap your own model
Install it, point it at your model, and measure the difference on your own long-horizon tasks.
Try it on Qwen 4B
A prepared build with the layer applied, so you can see whether it holds up before committing an afternoon.
$ pip install eunoia$ eunoia wrap --model <your-model>Measured on metrics that fit the problem.
No standard benchmark tells you whether a model is still obeying its instruction six hundred tokens in. These are the three we publish against, and the figures go up when they are measured rather than when they would be convenient.
Empty until measured. We would rather show an empty table than an unverifiable number.
What is not ready, stated plainly.
The rest of the platform is in progress. None of it is available, none of it has a date we are confident enough to publish, and we will not put a login form in front of something that cannot yet be logged into.
- Model APIs for Omega, Zeta and Sigma. None of these models have published results yet, so an API would be premature.
- Console and keys. Designed, not built.
- JS/TS and Python SDKs. Planned once the API surface stops moving.
- Status page and changelog. They arrive when there is something whose status is worth watching.
If you are building something serious.
Access to the larger models will roll out in phases, and the first phase will be small. If you are working on something where long-horizon reliability actually decides whether the product works, we would rather hear about it early and shape access around real use than guess.
Tell us what you are building and what breaks today. That is more useful to us than a waitlist entry.
What we publish, and what we keep.
Released
- The layer and its full source
- Configurations and evaluation benchmark
- The reference build and its base model
- Results, including negative ones
Retained
- CRSTO, our behaviour-aware compression research
- Adapter evolution logic
- Routing and control-loop heuristics
Open shell, private recipe. You can run it, change it, and put your own model under it without our internals.