TL;DR: NVIDIA released Nemotron 3 Nano, an open-weight language model that has 30 billion parameters on paper but only fires about 3 billion of them for any given token. It uses a hybrid Mamba-2 + Transformer mixture-of-experts design, carries a 1-million-token context window, and ships under a permissive commercial license with weights, datasets, and training recipes on Hugging Face. NVIDIA’s own numbers claim 3.3x the throughput of Qwen3-30B, and the third-party benchmark shop Artificial Analysis tracks it among the faster open models in its size class. The interesting part isn’t the leaderboard — it’s who’s building it, and why a chip company keeps giving models away.
The Story
Here’s a number that sounds like a contradiction: 30 billion parameters, 3 billion active. Nemotron 3 Nano is a 30B model — 31.6B if you want to be precise — but for any single token it passes through the network, it only lights up about 3.6B of those parameters. Out of 128 “experts” baked into the model, just 6 get called per forward pass.
That’s the mixture-of-experts (MoE) trick, and if the name scares you, think of it this way. A dense model runs every neuron on every word, whether the word needs them or not. An MoE model keeps a big roster of specialists and only pages in the handful relevant to the task at hand. You get the knowledge of a big model with the cost of a small one. Nemotron 3 Nano leans into that hard.
The architecture is the other half of the story. It’s a hybrid — interleaved Mamba-2 layers with grouped-query attention transformer layers. Mamba is a “state-space” design that scales more gracefully with sequence length than pure attention, which matters a lot when you’re advertising a 1-million-token context window. A million tokens is roughly a small library. For an agent, that means it can hold an entire task history, a pile of documents, and a multi-step plan in one window instead of chopping everything into chunks and hoping the retrieval stitches it back together.
NVIDIA is explicit about what this is for: agentic work. Software debugging, summarization, tool-use, information retrieval, long-horizon workflows where the model has to remember what it did twenty steps ago. It’s tuned to be cheap to run repeatedly, which is exactly the profile you want when an agent is calling the model in a loop hundreds of times to finish one job.
Now the numbers. NVIDIA’s own benchmarking claims Nemotron 3 Nano matches or beats Qwen3-30B and GPT-OSS-20B on accuracy while delivering 3.3x higher throughput than the former and 2.2x higher than the latter. It also claims 4x the throughput of its own predecessor, Nemotron 2 Nano, and up to 60% fewer “reasoning tokens” generated to reach an answer — which, since you pay per token, is a real cost lever. Those are first-party figures, so read them with the usual grain of salt. But there’s an outside check: Artificial Analysis, an independent benchmarking outfit, tracks Nano among the higher-throughput open models in its size class and scores it well on openness. When the vendor’s marketing and a third party point roughly the same direction, the claim is worth taking seriously.
Then there’s the word “open,” which gets abused constantly, so let’s be precise. Nemotron 3 Nano ships under the NVIDIA Nemotron Open Model License — a permissive, commercial-use-allowed license. NVIDIA didn’t just drop the weights on Hugging Face; it also published training datasets and the recipe, including the post-training stack (supervised fine-tuning, RLVR, RLHF), with some datasets gated behind an approval step. That’s meaningfully more open than a “you can download the weights, good luck reproducing it” release. It’s not the same as fully open-source in the classic sense — the license has terms, and not every dataset is unrestricted — so “open-weight with an open license and open recipe” is the honest description.
Nano is the small one. The same family also includes a Super tier (120B total, ~12.7B active) and an Ultra tier (550B total, ~55B active), sharing the architecture and the 1M-token context ambition. The weights live on Hugging Face and NVIDIA’s own build.nvidia.com endpoints, with the usual vLLM / TensorRT-LLM / SGLang deployment paths.
The Takeaway
I’ve been watching a very specific fight play out over the last stretch, and Nemotron 3 Nano is the latest move in it. Call it the “small-and-open efficiency race.”
Look at what’s landed recently. Liquid AI’s Nanos bet the future of agents on models tiny enough to run on your phone. Meta’s Muse Glimmer put a 30B agent model on a single GPU and reopened the open-weight fight. UNIST’s GMoE showed you could make neural layers share experts and cut a model by 63%. Different labs, different countries, one shared conviction: the frontier isn’t only about making models bigger, it’s about making capable models cheap to run. Nemotron 3 Nano is the same thesis with NVIDIA’s weight behind it — a 30B model that behaves like a 3B model at inference, given away open, tuned for agents.
And that’s the part worth sitting with. Why does a company that sells GPUs keep releasing free, hyper-efficient models? On the surface it looks like it’s undercutting its own hardware demand — an efficient model needs fewer chips. But that’s backwards. NVIDIA doesn’t make money selling you one inference. It makes money when AI gets deployed everywhere, running in loops, on edge devices, inside agents that call the model constantly. A cheap, open, agent-ready model is a demand generator: it lowers the barrier for building agentic systems, and those systems still run on someone’s silicon — ideally NVIDIA’s. The model is the razor; the compute is the blades. Shipping Cosmos 3 open and Nemotron 3 open isn’t charity. It’s seeding the ecosystem it profits from.
The architecture choice tells you where this is heading too. The pairing of Mamba with MoE is showing up again and again because it attacks the two costs that matter for agents: long context (Mamba handles length) and per-token compute (MoE handles width). We flagged the MoE efficiency angle when UNIST’s shared-expert work came through; seeing NVIDIA ship a production model on the same principle, with a million-token window, is the industrial version of that research bet paying off.
Here’s my read. The open-weight leaderboard isn’t the story. Anyone can top a benchmark for a week. The story is that “big model, small active footprint, run it anywhere, ship it open” has quietly become the default shape of a serious agent model in 2026 — and the company setting that default is the one that sells the compute underneath it. That’s a comfortable place to be. For everyone building on these models, the good news is real: capable open agents keep getting cheaper. The thing worth watching is whether “open license” stays genuinely open as the stakes climb, or whether the terms quietly tighten once the ecosystem is hooked.
This article is for informational purposes only and is not investment advice.
Photo: Igor Omilaev / Unsplash
댓글 남기기