PLaMo 2 1B

PLaMo 2 1B (pfnet/plamo-2-1b) is a 1-billion-parameter Japanese-focused language model from Preferred Networks' PLaMo 2 series, released in 2025 under Apache 2.0. It uses a hybrid Samba-style architecture that combines Mamba state-space layers with sliding-window attention, trained primarily on Japanese and English data totaling around 4 trillion tokens.

The PLaMo 2 technical report describes weight-reusing and pruning techniques that let PFN derive efficient small models from larger parents, and the 1B model achieves state-of-the-art Japanese results among similarly sized open models on instruction-following and fluency benchmarks. It is positioned for on-device and edge deployment where a fully Japanese-capable model must fit in tight memory budgets.