TII (Falcon)

A 0.6B image segmentation model.

Another small OCR model.

A series of small, hybrid models, going down to as little as 90M parameters.

Falcon is back! Similar to Nemotron-H, they use a hybrid architecture consisting of Attention and Mamba2 blocks for faster inference while needing less memory.

The Falcon series has also been updated. As a first among the "bigger" releases, they also release 1-bit models aimed at really fast inference.

Another strong open model in the Mamba architecture. No breakthrough characteristics, but Mistral is also playing in this space.

New model from Falcon team before Falcon 180 2, but not the best scores, weird license.