What is the model's architecture?

It uses a Mixture-of-Experts (MoE) structure with about 35 billion total parameters, but only about 3 billion are active during inference.

When will the model be available?

It will be added to the RICOH on-premises LLM Starter Kit from November.

Can it run on small servers?

Yes, an NVFP4 quantized version runs on a single NVIDIA DGX Spark small AI server.

Large Language ModelsAI Business & IndustryAI Watch (Impress)2h agoOct 15, 2026, 16:00 JST

Ricoh LMM_35B-A3B, 35B-parameter MoE, launches November

Send on LINE

3 Key Points

What happened
Ricoh developed "Ricoh LMM_35B-A3B," built on Alibaba Cloud's Qwen3.6-35B-A3B, with roughly 35 billion total parameters but only about 3 billion active at inference.
Why it matters
The model will be added to the RICOH on-premises LLM Starter Kit from November, giving enterprise users a Japanese-language option that runs efficiently.

WHO IT HITSEnterprise IT teams deploying on-premises AI models will gain a more resource-efficient Japanese-language option from November.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

Ricoh's new model builds on Alibaba Cloud's Qwen3.6-35B-A3B, combining it with Ricoh's own reinforcement learning and curriculum learning to boost Japanese reasoning. A key feature is its Mixture-of-Experts design, which activates only about 3 billion of its 35 billion parameters during inference, making it fast and resource-efficient.

Ricoh also developed an NVFP4 quantized version for NVIDIA Blackwell-generation GPUs. It runs on a single NVIDIA DGX Spark small AI server and is claimed to deliver reasoning performance exceeding the base model's quantized version while cutting response time per question to about one-third of the BF16 version (or about one-quarter with MTP). The model will be offered through the RICOH on-premises LLM Starter Kit starting in November.

FAQ

What is the model's architecture?: It uses a Mixture-of-Experts (MoE) structure with about 35 billion total parameters, but only about 3 billion are active during inference.
When will the model be available?: It will be added to the RICOH on-premises LLM Starter Kit from November.
Can it run on small servers?: Yes, an NVFP4 quantized version runs on a single NVIDIA DGX Spark small AI server.

AI Watch (Impress)Read Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.