Ricoh has developed the MoE-based multimodal large language model “RICOH LMM_35B-A3B.” The company aims to expand its lineup of on-premises LMMs that deliver improved Japanese chart and diagram comprehension with fast, resource-efficient inference.
Ricoh has developed “RICOH LMM_35B-A3B,” a mixture-of-experts (MoE) multimodal large language model designed to understand Japanese documents and charts with high accuracy and operate quickly with limited resources. Although it has approximately 35 billion total parameters, its architecture activates only approximately 3 billion parameters during inference.
The model is based on Alibaba Cloud’s Qwen3.6-35B-A3B and uses Ricoh’s proprietary reinforcement learning and curriculum learning to improve Japanese reasoning performance. It has outperformed open MoE models of a similar size on benchmarks including the JDocQA Reasoning Benchmark.
The model also supports Japanese AI agent functions, with improvements in evaluation metrics for its ability to automatically perform tasks using business systems and tools. In addition, Ricoh has strengthened its reward design to suppress repetitive looping responses, helping stabilize response times while maintaining output quality.
Ricoh has developed an NVFP4-quantized version that delivers practical generative AI processing on a single NVIDIA DGX Spark, a compact server designed to run on NVIDIA Blackwell-generation and later GPUs. Compared with the BF16 version, it significantly reduces response times per query.
The model is scheduled to be incorporated into the “RICOH On-Premises LLM Starter Kit” and made available through Ricoh Japan on a phased basis from November. By adding the model to a lineup that includes existing high-accuracy dense models and lightweight models, Ricoh will enable customers to select the best option for their requirements.