Ricoh Builds a 35B MoE Multimodal Model for On-Premises Japanese Document Reasoning

Ricoh has developed a multimodal large language model for companies that want generative AI to read Japanese documents, charts and tables without sending them to the cloud, ZDNET Japan reported on October 16. Announced a day earlier, Ricoh LMM_35B-A3B is built on Alibaba Cloud’s Qwen3.6-35B-A3B, a mixture-of-experts (MoE) model, and adds Ricoh’s own reinforcement-learning and curriculum-learning techniques to sharpen reasoning in Japanese.

The MoE design is the selling point. Of roughly 35 billion total parameters, only about 3 billion are active for any given token, because the model routes each input to a small set of specialist sub-networks. That keeps compute per token well below a dense model of similar size. The model handles contexts of up to 256,000 tokens and accepts video as well as images.

On Ricoh’s own JDocQA-Reasoning Benchmark v2.0, the score rose from 0.816 for the base model to 0.841, matching the dense Qwen3.6-27B while running about a ninth of the active parameters, and edging the commercial GPT-5.6 Sol at 0.821, according to the report. In a telecom customer-service scenario on the Japanese j-tau-bench agent benchmark, task success climbed from 62.4% to 77.7%, helped by suppressing looping answers.

What it means in practice

Ricoh Builds a 35B MoE Multimodal Model for On-Premises Japanese Document Reasoning — contextual photo

Ricoh also produced an NVFP4 quantised version for Nvidia Blackwell-generation GPUs that runs on a single DGX Spark box, cutting response time per question to about a third of the BF16 version. From November, Ricoh Japan will add the model to its RICOH on-premises LLM starter kit, alongside the dense Ricoh LMM_27B and the lightweight Ricoh LMM_9B released in June.

For buyers, the pitch lines up with wider moves to keep sensitive data inside Japan, such as NTT and SoftBank’s in-Japan AI cybersecurity that keeps data local. Evaluation is also getting stricter: AISI’s push for homegrown model benchmarks and the Sakura–PFN–NICT active evaluation platform could soon give procurement teams independent numbers to set beside vendor-run scores.

What to watch: whether third parties reproduce Ricoh’s benchmark gains, how public-sector buyers view a model derived from a Chinese open-weight base, and pricing for the starter kit when it ships.