NVIDIA releases Nemotron 3.5 Lightning, a fully open 30B mixture-of-experts

NVIDIA released Nemotron 3.5 Lightning on 11 August 2026, an open-weight mixture-of-experts model of roughly 30 billion parameters (reported as 31.6B) with about 3.6 billion active per token. It uses a hybrid Mamba-Transformer architecture, carries a one-million-token context window, and is positioned as an efficient workhorse for high-volume agentic deployments. NVIDIA describes it as distilled from its larger, unreleased Nemotron 3 Ultra, and shipped a companion router, NeMo Switchyard, that sends execution work to Lightning and harder planning to frontier models.

The licence is the part worth pausing on. Nemotron 3.5 Lightning is released under OpenMDW-1.1, which NVIDIA describes as permitting commercial use without material restrictions, and, unusually, it ships with the training data and training recipes alongside the weights. Most “open” models release only the weights. A model that also publishes its data and method is open in the fuller sense the research community actually asks for, because it can be reproduced and built on rather than only run. As always, read the licence text before you rely on it commercially.

Why it matters

Two things stand out. The first is who released it. NVIDIA sells the hardware most local AI runs on, so a genuinely open model from the GPU maker, data and recipes included, is a notable signal about where the company thinks the open ecosystem is heading. The second is the shape of the model. With only around 3.6 billion parameters active per token, it computes like a small model while holding the breadth of a 30-billion-parameter one, which is what makes it cheap to run at the high request volumes agents generate. NVIDIA says it runs across its range, from Jetson edge devices through the DGX Spark to data-centre systems.

For local use, a 30B mixture-of-experts is a familiar and practical size. As a rough guide it should need in the region of 18GB of VRAM at a 4-bit quantisation, comparable to other models in this class, though that is an estimate rather than an official figure and the hybrid architecture complicates it (see below).

Where it sits

Nemotron is NVIDIA’s open model family, and Lightning is the fast, distilled tier; it keeps the compact hybrid design of the earlier Nemotron 3 Nano while, NVIDIA says, improving intelligence and agentic performance. It arrives into a crowded field of small, fast open mixture-of-experts models. The obvious local comparison is Qwen3 30B-A3B, which is almost the same headline shape, a 30B MoE with roughly 3B active, and NVIDIA positions Lightning against models including gpt-oss-120b and the Qwen3.6 series. It also follows a run of open-model activity we have covered recently, from the Qwen3.6 mixture-of-experts to Zhipu’s GLM-5.3, whose own open weights are still pending.

The architecture, and the catch for local runners

The design is a hybrid: Mamba-2 layers, mixture-of-experts blocks, and attention combined. The Mamba parts are a different kind of sequence layer from the attention that dominates most models, and they are efficient over very long inputs, which is part of how the model reaches a one-million-token context and its high throughput. The catch is that this efficiency comes from an architecture that local inference engines support unevenly. Production GPU servers like vLLM and TensorRT-LLM are the surest route today; support for hybrid Mamba-Transformer models in the llama.cpp-based tools that most people run locally has been catching up, so check that your runtime handles this architecture before assuming a GGUF build will work.

On performance, treat the figures as claims until independent testing accumulates. NVIDIA’s own numbers, reported in BF16, include 51.56 on SWE-bench Verified, 75.44 on GPQA Diamond, and 81.94 on MMLU-Pro, along with claims of up to four times the output speed of similar-sized models and completing a 10,000-task run about 30% faster than Qwen3.6-35B at similar accuracy (NVIDIA, August 2026). The independent Artificial Analysis evaluation places it at 24 on its Intelligence Index, with an output speed near 670 tokens per second and 24% on Terminal-Bench v2.1 (Artificial Analysis, August 2026). The speed figures are the consistent theme across both, and speed at low cost is precisely the point for agentic workloads.

Worth a look if

You run models in the 30B mixture-of-experts class and want a fast, permissively licensed, genuinely open option for agentic or high-volume work, and you are on a runtime that supports the hybrid architecture. The full release of training data and recipes also makes it a more interesting starting point than most for anyone fine-tuning or researching, rather than only running, a model.