The post is dated 7 October 2026 and is authored by Pit Neitemeier and Alessio Serra.
The explorer lists Kolibri at 78.1B total parameters and 3.5B active parameters per token, or 4.4% active. At a 16K context it lists 28.3G FLOPs per token and 189M sequence-mixer state elements. Its displayed configuration includes 50 decoder layers, 384 routed experts, six routed experts active per token, four KV heads, and 40 sliding-window-attention layers with a window of 512.
For Kolibri Origin, the post lists 30.6B total and 3.3B active parameters. At one million tokens, it estimates 10.8B sequence-mixer state elements per Kolibri sequence, versus 53.7B for Origin. The article says Kolibri was recently released under Apache 2.0 and derives parameter, FLOP, and state calculations for MoE decoders.
The source does not establish model-quality gains, benchmark superiority, actual serving prices, or deployment performance for a particular hardware setup.