PLaMo 3.0 Prime

General availability of PFN's commercial flagship, in Reasoning (extended-RL "enhanced inference") and fast Non-reasoning variants. Context extended 64K→256K via YaRN plus continued pretraining; RL post-training roughly doubled versus the June 9 beta, with NICT safety data. PFN claims competitiveness with GPT-5.4 Mini, Claude Haiku 4.5, and Qwen3.6-27B on Japanese instruction-following and coding across 15 benchmarks (acknowledging weaker web-search, long-context, and math). API and on-prem, plus Bedrock Marketplace and Snowflake; no open weights; parameters and architecture undisclosed.