Nota, 2 MoE Quantization Papers Accepted at EMNLP 2026

This page was machine-translated and may differ from the original. View original

Nota, 2 MoE Quantization Papers Accepted at EMNLP 2026

Published2026.08.25 09:49

Qwen 3.8 Max GPU Requirements Reduced from 24 to 4 Units, LLM and Data Center Optimization Technology Application Expanding

Nota AI, an AI model lightweight and optimization company, has had 2 papers accepted at EMNLP 2026, a major international conference in natural language processing (NLP), continuing to deliver LLM optimization research results. The company is expanding the scope of technology application to the data center AI infrastructure sector, such as reducing the required NVIDIA B300 GPUs for running Qwen 3.8 Max from 24 to 4 units.

Nota announced on the 25th that papers on large-scale AI model optimization have been accepted at both the Main Conference and Findings of EMNLP 2026, with one paper in each category.

This year's conference received approximately 18,000 submissions, with a Main Conference acceptance rate of 15.4%.

EMNLP is a leading academic conference in the NLP field where companies such as Google, Meta, and OpenAI present their latest research.

Both accepted papers address the performance degradation issue that occurs in the quantization process of Mixture of Experts (MoE) models.

MoE performs computations using only some experts depending on the input, but the entire model must be loaded into memory, requiring large-scale GPUs for actual operation.

Reducing this through quantization can change expert selection, potentially degrading answer quality.

The Main Conference accepted paper presented the 'MENDS-MoE' methodology, which reflects the impact of quantization on next-step expert selection.

According to Nota, experiments with 4-bit and 3-bit quantization on 3 types of MoE models showed higher average accuracy compared to competing technologies.

The Findings accepted paper proposed the 'OPERA' methodology, which focuses optimization only on expert selection changes that actually affect responses.

Prior to this paper acceptance, Nota ranked 3rd among approximately 40 teams worldwide at the ICML 2026 'AdaptFM' workshop 'Efficient Qwen Competition'.

The company demonstrated running Qwen 3.5-4B on a single NVIDIA A10G GPU with inference speed 6.978 times faster on average compared to existing solutions.

Two MoE quantization papers were also accepted at the same workshop.

The company also disclosed actual model application results.

△Qwen 3.8 Max (over 1 trillion parameters) B300 GPU from 24 to 4 units △Moonshot AI 'Kimi K3' B300 GPU from 8 to maximum 4 units △Upstage 'Solar Open 2' H100 GPU from 8 to 2 units can each be reduced, Nota explained.

To date, Nota has published approximately 50 research results across major domestic and international conferences and journals, from core technologies such as pruning and quantization to vision language models (VLM) and LLMs.

Nota CEO Chae Myung-soo stated, "As the competitive axis in the AI industry shifts beyond model performance to operational efficiency, this achievement demonstrates that we have secured the technical foundation to improve the efficiency of large language models and data center AI infrastructure," and added, "We will continue to expand the scope of application of core optimization technologies that enhance the efficiency of AI infrastructure adoption."

#데이터센터 #AI 인프라 #양자화 #MoE #LLM 최적화 #Nota AI #Qwen #EMNLP 2026 #GPU 효율화

본 기사에 대한 정정·반론·추후보도 청구는 보도 청구 안내를, 그간 게재된 보도문은 정정·반론보도 모아보기를 참고해 주세요.

배종인 기자

All articles by this reporter

관련뉴스

Nota Launches AI Model Tuning Service Based on AWS Traynium and Inferentia

Nota, an AI model optimization technology company, has acquired AWS Advanced Partner certification and launched AI model tuning services based on Trainium and Inferentia. Through the Netspresso platform, the company provides one-stop support for technology validation, model porting, and performance tuning. The company has achieved a 68% reduction in model size and less than 1% accuracy loss on a language model with 32 billion parameters. This delivers practical cost efficiency to enterprises considering cloud AI chip adoption.

2026.07.07by 배종인 기자

NOTA Takes 3rd Place at ICML 2026 Challenge… Qwen Inference Optimization Technology Validated

AI model compression company Nota AI achieved 3rd place in the 'Efficient Qwen Competition' at the International Conference on Machine Learning ICML 2026. By combining quantization and speculative decoding techniques, the company improved the inference speed of Qwen3.5-4B by an average of 6.978x in a single A10G GPU environment. Additionally, two papers on MoE architecture LLM quantization were accepted at the AdaptFM workshop.

2026.07.13by 배종인 기자

Nota Joins AMD Robotics Partner Network… Expanding On-Device AI Robotics

Nota, a specialist in AI model optimization, has joined the AMD Robotics Partner Network to expand its on-device AI optimization technology into the robotics sector. The company plans to apply model optimization and hardware-tailored optimization techniques to reliably operate AI models in robotic environments where power, memory, and computational resources are limited.

2026.07.24by 배종인 기자

Nota Reduces AI Infrastructure Costs by Up to 76.5%... Unveils Lightweight Model Dedicated to 'Solar Open2'

Nota, a specialized AI model optimization company, has unveiled a lightweight version of Upstage's government-backed AI foundation model Solar Open2. By applying MoE quantization and pruning techniques, the company reduced weight memory by 76.5% and decreased the required GPUs from eight to two. The optimization is expected to deliver significant cost savings for enterprises' AI infrastructure expenses.

2026.07.27by 배종인 기자

노타, 문샷AI ‘키미 K3’ 경량화…엔비디아 B300 GPU 8장→4장

AI 모델 최적화 기업 노타가 문샷AI의 2조8,000억 매개변수 초거대 AI 모델 ‘키미 K3’를 경량화했다. 자체 개발한 비균일 전문가 프루닝 기술을 적용해 필요 GPU를 8장에서 4장으로 절감하면서도 핵심 성능을 유지했다. 이 기술은 층별 특성을 분석해 기여도 낮은 모듈만 선별 제거하며, 인프라 비용 절감을 기대할 수 있다.

2026.08.06by 배종인 기자