Eigen RadarAI
Archive
Analysis

Ricoh shrinks its Japanese AI safeguard model ahead of November rollout

Ricoh has developed a smaller Japanese model for checking harmful content in language-model inputs and outputs. Based on Gemma 4, it is intended to preserve the previous safety-classification level with fewer computing resources, according to the company. Customers can add labels for their own policies without further training. Phased inclusion in Ricoh’s on-premises language-model starter kit is scheduled from November.

Artificial Intelligence··Midday
An imagined compact ceramic sorter channels pale material through a radial core, with red-flecked pieces caught in side pockets.

A smaller model checks both sides of a conversation

Ricoh has developed a new Japanese safeguard model for detecting harmful information in the text sent to a large language model and in the responses it generates. The development, announced on October 20 and also covered by AI Watch, is intended for business use of generative AI. It evaluates supplied content and generated replies separately for safety. Ricoh plans to begin supplying the model to customers in November, through phased inclusion as a standard component of its on-premises language-model starter kit.[1], [2]

The new base is gemma-4-E4B-it, replacing Llama-3.1-Swallow-8B-Instruct-v0.5. Ricoh says the smaller design needs fewer computing resources while maintaining safety-classification performance comparable to its previous Llama-Ricoh-SafeGuard-20260520 model. The company links that reduction to use in on-premises installations with constrained computing capacity.[1]

Ricoh’s tests compare input and output classification

Ricoh compared the new model with its earlier Llama-Ricoh-SafeGuard-20260520 model. Its evaluation separately examines the text entering a language model and the responses leaving it, using distinct datasets for those two parts of the task. F1 combines precision and recall in classification. Input evaluation used 476 examples drawn from the National Institute of Informatics’ AnswerCarefully 2.0 dataset and Ricoh’s own data. Output evaluation used 508 examples prepared by Ricoh. The new input score is higher than the earlier model’s, while the output score is lower. These are separate classifications of material entering and leaving the language model.[1]

Companies can add labels without retraining

The update also revises safety-classification criteria. Ricoh says improvements in language models and broader practical uses have created more cases where purpose and context allow useful responses to content previously restricted uniformly. The stated aim is to reduce unnecessary limits. Alongside 13 pretrained general harmful-content categories, users can add labels reflecting their own operating policies, industry rules or company guidelines. Ricoh says that customization requires no additional model training. These features form part of the developed model scheduled for the November rollout.[1]

References

  1. News sourceRicohRicoh rebuilds its Japanese safeguard model on Gemma 4↩1↩2↩3↩4
  2. News sourceAI WatchRicoh switches its Japanese safeguard model to a Gemma 4 base↩
AI safety and regulationChip wars and the supply chainThe frontier model race

More analysis from this desk

View all Brief analyses from the Artificial Intelligence desk →