PLaMo 2

PLaMo 2 is Preferred Networks' second-generation family of Japanese-focused large language models, described in the September 2025 technical report. It uses a hybrid Samba-style architecture that transitions to full attention through continual pretraining to support 32K-token contexts. Training leans on extensive synthetic Japanese corpora to overcome data scarcity, with weight reuse and structured pruning for efficiency; the report says pruning produces an 8B model comparable in performance to PFN's previous 100B model. Post-training applies SFT and DPO on synthetic Japanese instruction data plus model merging, and the models are optimized for vLLM with quantization at minimal accuracy loss.

The flagship releases include PLaMo 2 1B, which is distributed under Apache 2.0, while larger variants ship under the PLaMo Community License, so check the license on the exact checkpoint you plan to use. PFN positions PLaMo 2 as state of the art among similarly sized open models on Japanese benchmarks for instruction following, fluency and Japan-specific knowledge. A related product, PLaMo Translation, specializes in Japanese-English translation for local and on-premise use.