Ricoh shrinks safeguard AI to gemma-4-E4B-it, ships in November

Ricoh's update centers on making its safeguard model lighter without sacrificing safety performance. By switching the base from Llama-3.1-Swallow-8B-Instruct-v0.5 to gemma-4-E4B-it, Ricoh says the new model matches the safety-detection level of Llama-Ricoh-SafeGuard-20260520 while requiring fewer computational resources. That matters for on-premise deployments, where customers run models on their own infrastructure and compute capacity is a practical constraint.

The revision also reflects how LLM use has spread. Ricoh notes that as LLMs have improved, content once uniformly restricted for safety reasons can now be answered usefully and appropriately depending on purpose and context. The new model therefore revisits the safety criteria to fit actual usage scenarios, aiming to balance safety with practicality.

Customization is another piece. Beyond the 13 pre-trained general harmful categories, users can add arbitrary judgment labels to match their own operational policies or industry rules, and this can be done without additional model training. Ricoh plans to standard-install the model in the RICOH オンプレLLMスターターキット from November and provide it to customers sequentially.