Kova Update-2 : The Big Leap
A look at the latest internal-beta build: browser use, TV control, day-to-day reliability, and a single request that brought several of those pieces together.
What this update means
Hello. Today we are sharing the latest Kova update. The past few weeks have been demanding: alongside our internal restructure, we have been working through DRM technologies and other restrictions that limited where and how Kova could help. We worked through those issues and are ready to show the progress from the newest build.
Kova remains in internal beta. That means the build is limited to internal developers and a small group of selected testers. It is not a public release, and the testing described here is meant to find rough edges before it becomes one.
Testing one request from start to finish
In the most recent internal-beta build, we tested browser use, TV use, reliability in day-to-day life, and a range of ease-of-use workflows. One of the most useful tests started with a single request:
“Hey, Kova, can you find me a good black cotton hoodie for work? Also, can you set a reminder for about 7 PM tomorrow that I need to mow the yard, and can you find me something to watch on the TV? Thanks, Kova.”
Kova worked through the request task by task. It handled the shopping search, created the reminder, and moved on to the TV request with good precision. Just as importantly, it correctly reported that no TVs were set up yet. That is the right answer: a helpful assistant should not pretend it can control a device that has not been connected.
Tests like this are useful because they are closer to how people actually ask for help. A real request may include a search, a reminder, and a device action in one sentence. We are measuring whether Kova can keep those threads straight, show its limits clearly, and make each step easy to review.
Watch the internal-beta build
You can also watch the Kova demonstration on YouTube.
The models we tested
This round of testing used Gemma 4 31B and our internal Deriva-0.6 model. Running the same workflows across both helps us understand what belongs to Kova's product and tool layer, and what depends on the model handling the request.
What still needs work
Voice interaction is still limited and error-prone, and TV control is not ready to rely on every day. Both remain active areas of work. We are focused on making voice interaction more dependable, improving TV setup and control, and continuing to test Kova against practical requests rather than only isolated demos.
There is more to do, but this build is a meaningful step forward. Stay tuned.