Eigen RadarAI
Archive
Analysis

Ricoh builds a sparse model for Japanese documents and tool use

Ricoh has developed LMM_35B-A3B, a model that activates only selected internal networks for each input. Based on Alibaba Cloud’s Qwen3.6, it targets Japanese business documents and agents that use tools while answering users. Ricoh plans to add the model to its locally hosted language-model starter kit beginning in November; its announced evaluations are company results.

Artificial Intelligence··Morning
A compact local computer and rear-facing monitor beside business documents.

Ricoh builds on Qwen for Japanese business documents

Ricoh developed LMM_35B-A3B, a multimodal language model based on Alibaba Cloud’s Qwen3.6-35B-A3B. The company applied additional training aimed at Japanese reasoning and business documents containing charts and diagrams.[1], [2]

The mixture-of-experts architecture divides processing among internal specialist networks and selects only those needed for each input. The model contains about 35 billion parameters, with about 3 billion active during inference. Ricoh used reinforcement learning and curriculum learning, which organizes the order of training tasks, in adapting the model.[1]

Agents are designed to use tools while handling requests

The model is intended to support AI agents that converse with users while operating business systems and tools. Ricoh examined that use in a Japanese telecom customer-service scenario. The training also includes rewards intended to discourage excessively long and repetitive answers while preserving response quality. The announced quality evaluations come from Ricoh’s own tests.[1]

A local deployment package is planned for November

Ricoh plans to begin adding the model to its locally hosted language-model starter kit in November. It also developed a compressed numerical version for NVIDIA Blackwell-generation and later graphics processors, designed for a single DGX Spark system. The planned package serves deployment in an organization’s own environment; the announcement does not describe a general release of the model weights.[1]

References

  1. News sourceAI WatchRicoh develops a sparse model for Japanese documents and tool use↩1↩2↩3↩4
  2. News sourceRicohRicoh develops LMM_35B-A3B for Japanese documents↩
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 →