A family of compact reasoning models, based off of the best 2B and 3B models, trained using improved DDP training code, no Unsloth.