이미지: X — 인프라·칩 화면 갈무리
Summary
- NVIDIA has released the open model Nemotron 3.5 Lightning, an MoE model with 30B total parameters and 3B active parameters
- The company says it delivers up to 4x faster output than similarly sized models
- The model went live for immediate use on Together AI the same day it was announced
- 모델명
- NVIDIA Nemotron 3.5 Lightning
- 구조
- 오픈 30B MoE, 활성 파라미터 3B
- 성능 주장
- 동급 크기 모델 대비 최대 4배 출력 속도(NVIDIA 발표)
- 타깃 용도
- 대량·특화 작업을 처리하는 상시 구동(always-on) 에이전트
- 제공처
- Together AI에 공개 당일 서비스 개시
- 발표 시점
- 2026년 8월 11일
Loading 30B, using only 3B
NVIDIA has released Nemotron 3.5 Lightning, the latest entry in its open Nemotron model lineup. The model has 300 billion total parameters, but only 3 billion of them actually participate in computation when generating a single token. The company says this architecture lets it produce answers up to 4 times faster than models of similar scale. Its target use cases are clear: agents that run around the clock without stopping, and workloads that repeat a fixed task at large volume.
On the same day, inference service provider Together AI added the model to its platform. Together AI described it as "the fastest open model in its class." The near-zero gap between the announcement and the service launch suggests this model wasn't designed as a research showcase, but as something meant to be plugged directly into an API right away.
Why 'active parameters' matter more than the headline number
Mixture of Experts (MoE) is an architecture that houses multiple sub-networks, called experts, inside a model, and for each input, only a subset of them is selected and used. It's like building an entire library but opening only the one or two shelves needed for a given question. The total amount of knowledge is at the 30B level, but computation cost and response latency approach the 3B level. This is why most open models released over the past year have converged on this structure. Qwen's Qwen-AgentWorld-35B-A3B, released on August 9, also had 35B total parameters with 3B active. A total in the 30B range with roughly 3B active — this configuration is hardening into something close to a standard in the open-model camp.
It's worth noting why speed has become so important. In a chatbot where a person asks once and gets one answer back, a one-second delay barely registers. But an agent calls the model dozens, sometimes hundreds, of times on its own to complete a single task. The latency of each call gets multiplied by that number of calls. This has created situations where it makes more sense to run a fast, cheap model many times, even at some cost to accuracy, and the name "Lightning" is aimed precisely at that point.
The fact that NVIDIA, a GPU company, keeps releasing models openly fits the same logic. Last July, the company signed an open letter supporting open weights alongside more than 200 companies and institutions, and has since released a string of open models, including the physical AI foundation model Cosmos 3 and Ising 1.5, a vision-language model for quantum computer calibration. The more models there are in the world capable of running on its hardware, the more chip demand follows.
So what changes
Teams trying to actually deploy agents in production now have one more option. The 30B/3B range sits in a middle ground: using a frontier model for everything is burdened by per-call cost and latency, while small models have struggled with task completion rates. The fact that teams can benchmark performance immediately through Together AI without needing their own GPUs is also practically significant. Details like benchmark scores and licensing terms will need verification once the model card is released, but it's clear this model sits squarely on the trend line shifting the center of gravity from "one big model" to "a fast model called many times."
