YuanLab AI Releases Yuan 3.0 Ultra: A Flagship Multimodal MoE Foundation Model, Built for Stronger Intelligence and Unrivaled Efficiency

2026-03-05 · Source: Machine Learning ML & Generative AI News · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Advanced, quick

Summary

YuanLab AI has released Yuan 3.0 Ultra, a trillion-parameter open-source multimodal Mixture-of-Experts (MoE) foundation model. This model achieves a 33.3% reduction in total parameters, from 1.5T to 1T, and a 49% increase in pre-training efficiency. These improvements are attributed to its novel Layer-Adaptive Expert Pruning (LAEP) algorithm, which prunes underutilized experts during pre-training. An Expert Rearranging algorithm further minimizes device-level token variance, leading to a high computational throughput of 92.6 TFLOPS per GPU. Yuan 3.0 Ultra also incorporates a Reflection Inhibition Reward Mechanism (RIRM) to promote concise reasoning, achieving leading accuracy on enterprise benchmarks like Docmatix (67.4%), ChatRAG (68.2%), and SummEval (62.8%).

Key takeaway

For AI Architects evaluating large-scale foundation models, Yuan 3.0 Ultra presents a compelling option due to its significant parameter reduction and enhanced pre-training efficiency. Its novel expert pruning and rearranging algorithms, combined with the RIRM, offer a path to more performant and concise reasoning models. Consider its benchmark performance and open-source nature for your next multimodal MoE deployment.

Key insights

Yuan 3.0 Ultra is a trillion-parameter MoE model optimized for efficiency and concise reasoning.

Principles

• Prune underutilized experts for efficiency.
• Minimize token variance for throughput.

Method

The model uses Layer-Adaptive Expert Pruning (LAEP) and Expert Rearranging algorithms for efficiency, alongside a Reflection Inhibition Reward Mechanism (RIRM) for concise reasoning.

In practice

• Achieves 92.6 TFLOPS per GPU.
• Reduces parameters by 33.3%.

Topics

• Yuan 3.0 Ultra
• Mixture-of-Experts
• Model Efficiency
• Multimodal Foundation Model
• Expert Pruning

Code references

• Yuan-lab-LLM/Yuan3.0-Ultra

Best for: AI Architect, NLP Engineer, Computer Vision Engineer, AI Engineer, Machine Learning Engineer, AI Researcher

Related on AIssential

• Qwen3.5-397B-A17B: the smallest Open-Opus class, very efficient model — Qwen3.5 offers multimodal, long-context capabilities with efficient sparse MoE architecture, enabling local deployment despite its large scale.
• The New Frontier: Benchmarking China’s Open-Weight Triad GLM 5.1, KIMI 2.6, MiMo-V2.5-Pro — Open-weight Chinese MoE models are outperforming proprietary Western AI on key developer benchmarks.
• Towards Faster Language Model Inference Using Mixture-of-Experts Flow Matching — MoE-FM enables faster, high-quality non-autoregressive language model inference by handling complex latent distributions.
• Latest open artifacts (#18): Arcee's 400B MoE, LiquidAI's underrated 1B model, new Kimi, and anticipation of a busy month — Smaller open models are achieving competitive performance against larger counterparts through extensive pretraining and architectural innovations.
• Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning — Nemotron 3 Ultra is an open-source hybrid MoE Mamba-Transformer achieving high throughput and accuracy for agentic AI.
• Mixture of Experts (MoE) - More Parameters, Same Compute — Mixture of Experts scales model capacity by having many specialized networks, but keeps per-token compute low by activating only a few.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning ML & Generative AI News.