PFN's new flagship in the PLaMo 3 line, released on 27 October 2026 as the successor to PLaMo 3.0 Prime. Like 3.0 Prime it is built on the pretrained model from PFN's joint research with NICT; the change is in post-training. Where 3.0 Prime shipped separate Reasoning and Non-reasoning variants, 3.1 Prime is a single hybrid model that switches between the two modes. The blog describes two-stage SFT (reasoning-only first, then a mix with reasoning traces stripped from a random subset), general RL for instruction following and STEM, a search-specific RL stage with rubric rewards for long-form reports (search simulated on an OpenSearch cluster), and DPO for the non-reasoning mode. New SFT data targets hallucination: answering known facts, admitting uncertainty, synthesized unanswerable questions, and citation data.

Serving changes: speculative decoding with a lightweight draft model gives a self-reported 1.70x output throughput at concurrency 1, falling to 1.28x at 32, and PFN's own low-bit KV-cache quantization cuts time per output token by 15–45%. Non-reasoning mode uses about 90% fewer output tokens. The product adds web search and Deep Research in PLaMo Chat, the API and on-premises deployments. Standard API pricing is ¥60 per million input tokens and ¥250 per million output tokens. Parameter count, architecture, context window and license are not disclosed (the evaluation ran at a 262,144-token max length in reasoning mode and 65,536 in non-reasoning mode). No weights are released.

Self-reported reasoning-mode scores, 3.0 Prime → 3.1 Prime: Japanese MT-Bench 93.1 → 97.3, AIME 2024 69.0 → 72.4, BrowseComp-Plus 17.8 → 21.8, AA-Omniscience 22.7 → 34.0, HealthBench 16.2 → 19.4. Some scores fell: IFBench 72.2 → 64.8, LiveCodeBench 48.3 → 44.7, GPQA-Diamond 62.2 → 61.2. Non-reasoning mode gave up most of its maths (AIME 2024 58.0 → 10.2, MATH-500 91.1 → 63.5). PFN reports an average gain of about 15% over 3.0 Prime across eight hallucination benchmarks. Against gpt-oss-120b, Gemma 4 31B IT, Qwen3.8 27B and Muse Glimmer 30B, the blog says it leads on safety in both modes and trails all of them on STEM/code and web-search QA. Artificial Analysis does not list PLaMo models.

Model Details

Benchmark Scores

Benchmark Score Mode
AIME 2024 72.4 reasoning
GPQA-Diamond 61.2 reasoning
IFBench 64.8 reasoning
LiveCodeBench 44.7 reasoning
Japanese MT-Bench 97.3 reasoning
AA-Omniscience 34.0 reasoning
japaneseproprietaryreasoningsovereign

Related