All research
02 / Model
Kestrel: open models, in the browser or on a server
Kestrel Lite is Qwen3.5 0.8B in your browser; Kestrel Pro is Gemma 4 26B on a server. Avenpa has not fine-tuned either.
The Kestrel line uses other groups' open models instead of training from scratch. Kestrel Lite is Qwen3.5 0.8B (Apache-2.0), small enough to load quickly in a browser. Kestrel Pro is Google's Gemma 4 26B, which is too large for a browser and runs on Cloudflare's servers.
Neither is fine-tuned by Avenpa. Kestrel Pro needs a signed-in account and sends your messages to the server; Kestrel Lite runs entirely on your device. Korean fine-tuning is planned.