Nemotron 3.5 Lightning: One GPU, Yes. A Laptop, No.

NVIDIA's 30B Nemotron 3.5 Lightning runs on a single desktop GPU, but its Q4_K_M GGUF is 25.3 GB — above every 24 GB mobile card.

by HowAIWorks Team
On this page

Introduction

NVIDIA released Nemotron 3.5 Lightning on 11 August 2026: a 30-billion-parameter Mixture-of-Experts model with roughly 3B active parameters, a 1M-token context window, and an OpenMDW-1.1 licence. The coverage settled quickly on one line — it runs locally, on hardware you already own.

That line is true in a narrower sense than most of the write-ups implied, and the narrower version is the more interesting story. NVIDIA never said "laptop." Its named local targets are RTX PCs, RTX PRO workstations, DGX Spark, DGX Station and Jetson. The default 4-bit GGUF is 25.3 GB, which is above the ceiling of every mobile GPU shipping today. What NVIDIA actually built is a model that fits on one desktop-class card — and that is a real achievement that does not need to be inflated into a laptop claim.

The timing makes the point twice over. The day before, on 10 August, Meta released Muse Glimmer — also 30B, also "runs locally," also with a headline VRAM number that only holds for one specific quantized build. Two open 30B models, one day apart, with the same gap between the marketing and the file sizes.

What NVIDIA actually shipped

Lightning is the successor to Nemotron 3 Nano, not a small sibling of Nemotron 3 Ultra. It is a hybrid Mamba-2 / MoE / attention stack, pre-trained on more than 20 trillion tokens with an NVFP4 recipe, with a pre-training data cutoff of September 2025 and a post-training cutoff of May 2026.

Three things about it are unusual for a model this size:

  • The context ceiling. 1M tokens, the same figure NVIDIA quotes for Ultra at 550B parameters.
  • The activation ratio. ~3B active out of 30B means a forward pass costs about what a 3B dense model costs, while the parameter count buys the knowledge of something much larger.
  • The licence. OpenMDW-1.1, the Linux Foundation's permissive model licence, covering weights, recipe and redistributable data — the same terms as Ultra, and more permissive than the NVIDIA Open Model License that Nano and Super shipped under.

One correction worth making early, because it is widely assumed: there is no Nemotron 3.5 size ladder. No 3.5 Super, no 3.5 Ultra. Lightning is the only general-purpose model carrying the 3.5 label. A separately-scoped Nemotron 3.5 Content Safety guard model exists — 4B, a Gemma 3 4B fine-tune under the NVIDIA Open Model License, released 4 June 2026 — but it shares the version string and nothing else.

The hardware list has no laptop on it

NVIDIA's own materials are precise, and the imprecision came downstream. The developer blog names "NVIDIA Jetson, GeForce RTX 5090, and DGX Spark." The company blog widens that to RTX PCs, DGX Spark, DGX Station and Jetson locally, scaling out to edge devices, RTX PRO workstations, data centres and cloud. Every one of those is a desktop, a workstation, a dev box or an embedded board.

The validated single-GPU deployments on the BF16 model card are equally specific:

ConfigurationNotes
1x H100 80GBBF16, memory-bound at ~256K context
1x A100 80GBvia W4A16 quantization
1x GB200 / 1x B200native FP4 tensor cores
DGX Spark (GB10)validated at the full 1M context
8x H100 (TP8 + expert parallel)the path to 1M on Hopper

Read the first row again. A single H100 — an 80 GB data-centre accelerator — tops out around a quarter of the advertised context in full precision. The 1M-token headline is a property of the architecture, not of any one machine you are likely to have.

The GGUF is 25.3 GB

For local runtimes the number that decides everything is the file size, and Unsloth's GGUF conversion settles it:

BuildSize
BF1665.9 GB
Q8_035 GB
Q5_K_M30.2 GB
Q4_K_XL25.5 GB
Q4_K_M (default)25.3 GB
Q4_K_S24.5 GB
Q3_K_XL21.2 GB

The largest mobile GeForce card on sale carries 24 GB. The default Q4_K_M does not fit it. Neither does Q4_K_S, at 24.5 GB, before a single token of KV cache is allocated. A 24 GB laptop runs Lightning at Q3_K_XL or not at all — which is a legitimate option, and also a materially more quantized model than the one NVIDIA benchmarked.

On a 32 GB desktop card — the RTX 5090, NVIDIA's named target — Q4_K_M leaves roughly 7 GB for context and overhead. That is the configuration the "runs on one GPU" claim describes, and it works. It is a desktop claim.

Muse Glimmer has the same shape of gap

Meta's Muse Glimmer arrived on 10 August under Apache 2.0 with an explicit 24 GB promise, and unlike Lightning's, that promise holds — but only just, and only for one build.

The K-Quant-17GB (Q4_K_M) file is 16.8 GB and wants about 17 GB of VRAM for text. Then the parts that make it interesting get loaded. The ~1.8B perception encoder adds about 1.4 GB, taking you to ~19 GB. The DFlash speculative decoding drafter — the thing that makes a dense 30B model tolerable in an agent loop — adds ~1.6 GB, taking you to ~20 GB before any context is cached. On a 24 GB RTX 3090 or 4090 that leaves roughly 4 GB for a window Meta advertises at 128K.

So the two models fail the same way from opposite directions:

  • Lightning is sparse, fast and long-context, and its default quant simply does not fit 24 GB.
  • Glimmer fits 24 GB, but only with the vision encoder and drafter unloaded — and those are the two features you downloaded it for.

Neither is a scandal. Both are what happens when a headline is written against a best-case build and read as a description of the model.

The score that needs a version number

There is a second number worth handling carefully. Artificial Analysis has published two Intelligence Index scores for Lightning, and both are correct.

Its launch write-up on 11 August 2026 scored Lightning 24 on Intelligence Index v4.1, the version current from 15 June 2026 — a nine-point gain over Nemotron 3 Nano's 15 on the same index, and roughly level with gpt-oss-120b at about a quarter of the total parameters. Its model page today shows 16 on Intelligence Index v4.2, which shipped on 4 September 2026, added AA-Briefcase and GDP.pdf, retired a saturated GPQA Diamond and doubled private held-out weighting to 40%.

The v4.1 and v4.2 numbers are not comparable, and the drop from 24 to 16 is a change of ruler, not of model. This is the whole reason a score without its index version is not a fact — and Lightning is not alone: on the same transition Grok 4.6 went from 61 to 51 and Muse Spark 1.3 from 61 to 53.

What Lightning is actually for

Strip the local-hardware framing and the positioning is coherent. Lightning is an execution model for agents: tool calls, argument construction, result validation, sub-agent delegation — the high-volume steps between the reasoning, where latency is paid a thousand times per run and a frontier model adds cost to a decision that was never hard.

NVIDIA's speed claims fit that framing, and so does their most honest detail: up to 4x the output speed of similar-sized models, but 10,000 PinchBench tasks completed only 30% faster than Qwen3.6 35B. A 4x token speedup buying a 30% task speedup is a clean measurement of where the time actually goes in an agent loop — orchestration and tool latency, not generation.

On OpenRouter, Lightning serves at about $0.08 / $0.20 per million tokens with 262K context and 131K max output; Nemotron 3 Ultra sits at $0.50 / $2.20 with a 16K output cap. That six-to-one price gap is the reason NVIDIA shipped NeMo Switchyard alongside it — a router that sends plans up to a frontier model and execution down to Lightning. Locally, Ollama, LM Studio and llama.cpp all consume the GGUF builds.

Worth watching separately: the Nemotron Coalition, announced 16 March 2026 with Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab. Its first output is a base model co-developed by Mistral AI and NVIDIA, which NVIDIA says will underpin the coming Nemotron 4 family. NVIDIA's next open generation will not be purely NVIDIA-built — an unusual position for a company that sells the hardware everyone else trains on.

Conclusion

Nemotron 3.5 Lightning is a good model with an overstated distribution story. A 30B MoE with ~3B active parameters, a million-token ceiling and a permissive licence, running on a single desktop GPU, is a genuinely useful thing to be able to download — and 25.3 GB is the number that tells you which desk it fits on.

The practical rule for both of the August 30B releases is the same: treat every "runs locally" headline as a claim about one quantized build, then go read the file sizes. Lightning needs 32 GB for its default quant. Glimmer fits 24 GB only with its vision encoder and drafter unloaded. And any Intelligence Index score you quote needs its index version attached — v4.1 and v4.2 are different rulers — or it is not telling you what you think it is.

Sources

Frequently Asked Questions

Not at the default quantization. The Unsloth Q4_K_M GGUF is 25.3 GB and the Q4_K_S is 24.5 GB, both above the 24 GB ceiling of the largest mobile GeForce card. NVIDIA's own named local targets are RTX PCs, RTX PRO workstations, DGX Spark, DGX Station and Jetson — desktop and embedded hardware, no laptop. A 24 GB laptop can run the smaller Q3_K_XL build at 21.2 GB, with the accuracy cost that implies.
It depends entirely on the build. NVIDIA validates single-GPU deployment on 1x H100 80GB, 1x A100 80GB (via W4A16), 1x GB200 or B200, and DGX Spark. For desktop use the GeForce RTX 5090 is the named target, and the community GGUF quants run from 21.2 GB (Q3_K_XL) through 25.3 GB (Q4_K_M) to 65.9 GB at BF16. A 32 GB card is the comfortable floor for the default 4-bit build.
Because they were measured on two different indexes. Artificial Analysis scored Lightning 24 in its launch write-up on 11 August 2026, when Intelligence Index v4.1 was current; its model page now shows 16 on Intelligence Index v4.2, which shipped on 4 September 2026. v4.2 retired GPQA Diamond, added AA-Briefcase and GDP.pdf, and doubled private held-out weighting. The two numbers are not comparable and neither is a correction of the other.
They are the same size and one day apart, but architecturally opposite. Lightning is a 30B sparse Mixture-of-Experts with roughly 3B active parameters and a 1M-token context; Muse Glimmer is a dense ~29.6B model with a 128K context and a vision encoder. Glimmer's 4-bit build genuinely fits 24 GB, Lightning's does not. Lightning is faster per token; Glimmer reads images.
About 256K tokens. NVIDIA's BF16 model card marks the single-H100 configuration as memory-bound at that length, well short of the model's 1M-token ceiling. Reaching the full million tokens on one device means the NVFP4 build or a DGX Spark; on H100 it takes eight GPUs with tensor and expert parallelism.

Continue Your AI Journey

Explore our glossary and model catalog to deepen your understanding.