Nemotron Runs 30B on One Consumer GPU
Two 30B open agentic models, twenty-four hours apart. That's the story, not either one alone.
Founder, Automation Squad ·
The short answer
NVIDIA released Nemotron 3.5 Lightning on August 11, 2026 — a 30B open Mixture-of-Experts model with only 3B active parameters per token, distilled from Nemotron 3 Ultra. It runs on a single consumer GPU, is licensed under OpenMDW-1.1 with weights, training data and recipes released, and claims up to 4× the output speed of similar-sized models.
Nemotron 3.5 Lightning against the other open 30B that shipped this week
Two open agentic models at the same parameter count, a day apart, with different architectures and different licences. Worth comparing directly.
Understand what 3B active parameters actually buys you
A Mixture-of-Experts model holds 30B parameters but only runs 3B of them per token. You need memory for the full model and compute for a fraction of it — which is why speed improves without the quality drop of simply using a smaller model.
Check your GPU memory before your GPU speed
MoE shifts the constraint to memory. Both of this week's releases are sized for a single consumer card, but the sizing assumptions differ, so read the specific numbers rather than the headline parameter count.
Read the licences — they are genuinely different
Muse Glimmer is Apache 2.0, which is permissive and familiar. Nemotron 3.5 Lightning is OpenMDW-1.1. If you are shipping something commercially, that difference deserves five minutes with whoever owns legal.
Look at NeMo Switchyard even if you skip the model
NVIDIA shipped an open routing library that sends each task to the cheapest model capable of handling it. That pattern transfers directly to n8n, LangGraph or anything else you orchestrate with, regardless of which models you run.
Test on your job, not on their benchmark
Both vendors published favourable comparisons against different competitors. Run fifty of your real inputs through each and count disagreements — that number is worth more than any published benchmark.
| Nemotron 3.5 Lightning | Muse Glimmer | |
|---|---|---|
| Released | Aug 11, 2026 (NVIDIA) | Aug 10, 2026 (Meta) |
| Parameters | 30B MoE, 3B active per token | 30B dense, multimodal |
| Licence | OpenMDW-1.1 | Apache 2.0 |
| Released alongside weights | Training data and recipes | Weights + GGUF |
| Runs on | Single consumer GPU (RTX, DGX Spark) | Under 20GB at 4-bit; 24-32GB envelope |
| Claimed speed | Up to 4× similar-sized models | Ships with DFlash speculative decoding |
| Benchmark claim | 86% PinchBench, 10,000 tasks 30% faster than Qwen3.6 35B | Benchmarked vs Gemma4-31B, Qwen3.6-27B |
| Shipped with | NeMo Switchyard routing library | vLLM, SGLang, Ollama, LM Studio |
Take it with you
TWO OPEN 30B AGENTIC MODELS, 24 HOURS APART
Source: https://automationsquad.com/news/nvidia-nemotron-lightning/
NEMOTRON 3.5 LIGHTNING MUSE GLIMMER
Released ....... Aug 11 (NVIDIA) Aug 10 (Meta)
Params ......... 30B MoE, 3B active/token 30B dense, multimodal
Licence ........ OpenMDW-1.1 Apache 2.0
Also released .. training data + recipes weights + GGUF
Runs on ........ single consumer GPU <20GB at 4-bit
Speed claim .... up to 4x similar sizes ships DFlash spec decoding
Shipped with ... NeMo Switchyard router vLLM/SGLang/Ollama/LM Studio
WHAT "3B ACTIVE" MEANS
Holds 30B parameters, runs 3B per token. You need MEMORY for the full
model, COMPUTE for a fraction. Speed without the quality drop of just
using a smaller model.
CHECKLIST
[ ] Check GPU MEMORY before GPU speed (MoE shifts the constraint)
[ ] Read both licences — Apache 2.0 vs OpenMDW-1.1 is a real difference
if you're shipping commercially
[ ] Look at NeMo Switchyard even if you skip the model — routing each
task to the cheapest capable model transfers to n8n/LangGraph
[ ] Run 50 of MY real inputs through each. Count disagreements.
That beats any published benchmark.NVIDIA released Nemotron 3.5 Lightning on August 11, 2026 — a 30B open Mixture-of-Experts model with roughly 3B active parameters per token, distilled from Nemotron 3 Ultra. It arrived a day after Meta released its own 30B open agentic model.
The facts: NVIDIA claims up to 4× the output speed of similar-sized models, 86% accuracy on PinchBench, and completion of 10,000 tasks 30% faster than Qwen3.6 35B. It runs on a single consumer GPU including RTX PCs and DGX Spark. The licence is OpenMDW-1.1, and weights, training data and recipes were all released. It shipped alongside NeMo Switchyard, an open routing library that directs each task to the cheapest model capable of handling it.
Automation Squad's take: the individual release is good and the coincidence is more interesting. Two open 30B agentic models from two of the largest companies in the industry, twenty-four hours apart, both sized deliberately to run on hardware an individual already owns. That is not a coincidence about timing, it is a signal about where both companies think the useful frontier is — and it lines up exactly with the download data showing that small models are what people actually run. The Mixture-of-Experts approach is the more durable idea here: holding 30B and running 3B is a better trade than dropping to a 3B model outright, and it is the architecture doing that work rather than the branding.
Run this now: check your GPU's memory rather than its speed, because Mixture-of-Experts moves the constraint. Then pick the job you currently send to a large hosted model that involves no confidential reasoning and high volume, and try it locally on one of this week's two releases. Count disagreements across fifty real inputs. And read both licences before you build anything commercial on either — Apache 2.0 and OpenMDW-1.1 are not interchangeable.
Questions people are asking
- What does 30B with 3B active parameters mean?
- It is a Mixture-of-Experts architecture. The model contains 30 billion parameters but activates only about 3 billion for any given token, so it needs memory sized for the full model and compute sized for a fraction — which is how it gets speed without the quality cost of simply using a smaller model.
- Can I use Nemotron 3.5 Lightning commercially?
- NVIDIA released it under OpenMDW-1.1 with weights, training data and recipes published. That is a different licence from Meta's Apache 2.0 release the previous day, so if you are shipping commercially the two are worth comparing properly rather than treating both as simply open.
- What is NeMo Switchyard?
- An open routing library NVIDIA shipped alongside the model that sends each task to the cheapest model capable of handling it. The pattern is reusable in any orchestration setup, independent of which models you run.
Sources
Further reading
- NVIDIA — Nemotron 3.5 Lightning — Full benchmarks, the licence terms, and the NeMo Switchyard routing library.
Last checked August 11, 2026 against the primary sources above, by Automation Squad Research. Spot an error? [email protected].