Inspiration

Small models are fast and efficient, while large models offer stronger reasoning at higher computational cost. We asked: How can a model know when it needs a bigger brain?

What it does

Horizon Router starts with K2 Horizon 0.9B and evaluates whether its response can be trusted. Easy tasks stay local; difficult or uncertain tasks are escalated to K2 Horizon 375B.

How we built it

K2 Horizon 0.9B first attempts the task. We then evaluate signals such as uncertainty, consistency, critic feedback, and task complexity to generate an escalation score and determine whether to keep the response or route it to a larger model.

Challenges we ran into

The biggest challenge was determining when an answer is truly unreliable. Confidence alone isn't enough, so we combined multiple signals while balancing unnecessary escalations against missed difficult cases.

Accomplishments that we're proud of

We built a working end-to-end adaptive routing system in one day that decides when additional model capability is needed.

What we learned

Model routing is itself a reasoning problem. Task difficulty isn't always obvious from the prompt, and combining multiple signals provides a better indication of when escalation is necessary.

What's next for Horizon Router

We'd like to test on larger benchmarks, improve threshold calibration, and eventually extend Horizon Router to dynamically route across multiple models and capability levels.

Built With

Share this project:

Updates

Submission history