PFN's PLaMo 3.1 Prime puts reasoning and fast answers in one model, so serving no longer needs double the compute

One model now covers both reasoning and non-reasoning modes: PLaMo 3.1 Prime follows the mode named in each request when deciding whether to write out a reasoning trace before answering. PLaMo 3.0 Prime built the two modes as two models, with separate instruction-tuning and reinforcement-learning pipelines; under the configuration reported at the time, offering both took twice the compute of serving one single-mode model such as PLaMo 2.2 Prime or PLaMo 3.0 Prime Beta. The comparison covers 35 benchmarks in nine categories: both modes score above PLaMo 3.0 Prime in most categories, the reasoning-mode web-search report category averages 11.1 points higher, and safety is the highest among the compared open models in both modes.

The two modes are taught to coexist by mixing training data, not by keeping a separate model for each mode. Instruction tuning runs in two stages: the first uses only data carrying reasoning traces, the second strips the traces from a fraction of examples and mixes them back; general reinforcement learning and reinforcement learning on web-search tasks follow, and finally DPO, which tunes outputs against paired better and worse answers, adjusts responses centered on non-reasoning-mode tasks. Scores recorded at each training stage show that after the mixed stage, continued training on reasoning data alone still lifts the non-reasoning mode alongside it; STEM/Code non-reasoning scores fall, because the second-stage instruction tuning deliberately suppresses the tendency of non-reasoning replies to grow long and trains faster answers.

Serving speed comes from two engineering changes: speculative decoding has a lightweight draft model guess several tokens for the main model to verify in one pass, and low-bit KV cache quantization stores the key-value cache of processed tokens at lower precision. An LLM emits one token at a time and re-reads its weights from GPU memory for each, so bandwidth rather than arithmetic is the bottleneck; the draft model's matching tokens are accepted, and mismatches are resampled from the main model's distribution, so quality matches decoding without a draft. On NVIDIA's SPEED-Bench the output throughput ratio is 1.70x at concurrency 1 and falls to 1.28x at concurrency 32, since busier GPUs have less spare compute to absorb verification; KV cache quantization shortens per-token generation time by 15-45% at the same batch size.

Evaluation ran on the company's own GPU cluster with vLLM, covering 35 benchmarks grouped into nine categories, and the model is served through PLaMo Chat and the API; production deployment is optimized for the target hardware, so behavior and performance may differ from the evaluation. The stated next step is to make the amount of reasoning adjustable so users can trade tokens against performance by use case, and to keep strengthening web-search QA and long-context RAG while building coding-agent capability.

Two regressions are reported explicitly: instruction following in reasoning mode and STEM/Code in non-reasoning mode score below PLaMo 3.0 Prime. Against the open models in the same performance band, STEM/Code and web-search QA trail in both modes, and dialogue and long-context RAG are often weaker; the non-reasoning mode uses the fewest tokens of the compared models but scores well below Gemma 4 31B IT and Qwen3.8 27B in their non-reasoning modes.