Kova Update-1: internal beta begins.

Good enough for daily use, not ready for release

Today we're marking a real milestone: Kova has moved into internal beta. That means it's good enough for us to use day-to-day, not just to test in isolated demos. It does not mean Kova is release-ready. It is still very rough around the edges, and there is a long list of polish, stability, and safety work between here and a public launch. But reaching daily-use quality internally is proof the core approach works, and it has opened the door to testing capabilities we could not responsibly test before.

Why we're testing under tighter restrictions than we'll ship with

Kova is powerful, and power without care around it is a liability. It can act with real consequences: spend money, control your home, browse the web, write and run code. Because of that, we are deliberately testing it under more restrictive settings internally than we plan to ship publicly. That sounds backwards, but it isn't: it lets us find where the restrictions actually matter before we decide how to relax them, so the version people eventually get is capable and safe by default, not safe because it's crippled.

The models behind this round of testing

Internal beta testing has run across a wide mix of models, not just our own, so we can see how Kova performs regardless of what's doing the reasoning underneath. So far, that list includes:

• Deriva-0.2 (non-release)
• Hikaru-0.1 (non-release)
• MiniMax-M3
• Claude Opus 5
• Fable 5
• GPT-5.6 SOL
• Ornith-1.5:9b (via the developer integration)
• Grok-4.6
• Llama 3.1
• Several additional internal models we can't disclose yet

We went into this round expecting Kova to be usable. What surprised us was how far it could be pushed, even under restriction, across such a wide range of models.

What comes next

Internal beta continues, and later updates will cover what changes as we work through stability, safety, and the rest of the road to a public release. Stay tuned.