NVIDIANemotronNemotronHForCausalLM

NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 — Hardware Requirements & GPU Compatibility

ChatFunctions

NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 is a mixture-of-experts chat model in NVIDIA's Nemotron line, with roughly 30 billion total parameters and about 3 billion active per token. That active-parameter count drives inference speed, while the larger total governs memory footprint, and the Lightning name points to a variant tuned for fast responses; weights are published in BF16 precision. The model offers a 256K token context window and is released under a custom license rather than a standard open-source one. With only about 3 billion parameters active per token, it is well suited to running on a single high-end consumer GPU, especially once quantized, despite its larger total parameter count.

31.6B params · MoE2.4B active262K contextOtherAug 2026689.1K downloads238 likes1.3M quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Runs great on a 24 GB GPU

Runs from 12 GB (IQ2_M · 128K)Best on 24 GB (Q4_K_M · ~197 tok/s · 256K)Macs from 16 GB (IQ2_M · ~61 tok/s)

Other

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on Tesla P40 (24 GB) (4 GB headroom) · Q4_K_M

Generation speed: ~108 tok/s; generation speed
Cost per 1M output tokens: $0.31; per 1M output tokens

Deployon Vast.aiBest optionfrom $0.12/hr

Compare GPUs →

Run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16

19.3 GB VRAM at 16K context · fits the RTX 4090 (4.7 GB free)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~69 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model TensorFold/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-MLX-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$unsloth/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
Tesla P40 (24 GB)CheapestCheapest4 GB free	~108 tok/s	$0.31	Deploy on Vast.aifrom $0.12/hr (opens in new tab)	24 GB · 4 GB free · ~108 tok/s · $0.31 / 1M tokens
RTX A5000 (24 GB)4 GB free	~174 tok/s	$0.26	Deploy on RunPodfrom $0.16/hr (opens in new tab)	24 GB · 4 GB free · ~174 tok/s · $0.26 / 1M tokens
RTX A5000 (24 GB)4 GB free	~174 tok/s	$0.34	Deploy on Vast.aifrom $0.21/hr (opens in new tab)	24 GB · 4 GB free · ~174 tok/s · $0.34 / 1M tokens
RTX 3090 (24 GB)4 GB free	~191 tok/s	$0.32	Deploy on RunPodfrom $0.22/hr (opens in new tab)	24 GB · 4 GB free · ~191 tok/s · $0.32 / 1M tokens
RTX 3090 (24 GB)4 GB free	~191 tok/s	$0.37	Deploy on Vast.aifrom $0.26/hr (opens in new tab)	24 GB · 4 GB free · ~191 tok/s · $0.37 / 1M tokens
RTX PRO 4000 (24 GB)4 GB free	~162 tok/s	$0.58	Deploy on Vast.aifrom $0.34/hr (opens in new tab)	24 GB · 4 GB free · ~162 tok/s · $0.58 / 1M tokens
RTX PRO 4500 (32 GB)12 GB free	~187 tok/s	$0.50	Deploy on RunPodfrom $0.34/hr (opens in new tab)	32 GB · 12 GB free · ~187 tok/s · $0.50 / 1M tokens
RTX 4090 (24 GB)4 GB free	~197 tok/s	$0.48	Deploy on RunPodfrom $0.34/hr (opens in new tab)	24 GB · 4 GB free · ~197 tok/s · $0.48 / 1M tokens

Compare GPUs →

How Much VRAM Does NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
9.1 GB	9.1 GB	9.6 GB	~8.7 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 256K context · ~257 tok/s
9.8 GB	9.9 GB	10.4 GB	~9.5 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 256K context · ~251 tok/s
10.2 GB	10.3 GB	10.8 GB	~9.9 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 256K context · ~248 tok/s
11.0 GB	11.1 GB	11.6 GB	~10.7 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 256K context · ~243 tok/s
12.6 GB	12.7 GB	13.2 GB	~12.2 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 256K context · ~233 tok/s
13.4 GB	13.5 GB	13.9 GB	~13 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 256K context · ~228 tok/s
13.8 GB	13.9 GB	14.3 GB	~13.4 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 256K context · ~226 tok/s
13.8 GB	13.9 GB	14.3 GB	18.9 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 256K context · ~226 tok/s
14.2 GB	14.3 GB	14.7 GB	18.9 GB	3-bit small — 3-bit small quantization	Fits: 256K context · ~223 tok/s
14.6 GB	14.7 GB	15.1 GB	20.3 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 256K context · ~221 tok/s
15.8 GB	15.9 GB	16.3 GB	19.8 GB	3-bit medium — 3-bit medium quantization	Fits: 256K context · ~215 tok/s
16.2 GB	16.2 GB	16.7 GB	~15.8 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 256K context · ~213 tok/s
16.6 GB	16.6 GB	17.1 GB	20.2 GB	3-bit large — 3-bit large quantization	Fits: 256K context · ~211 tok/s
17.3 GB	17.4 GB	17.9 GB	18.9 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 256K context · ~207 tok/s
18.1 GB	18.2 GB	18.7 GB	21.2 GB	4-bit · i-quant non-linear — Importance-weighted 4-bit, non-linear	Fits: 256K context · ~203 tok/s
18.1 GB	18.2 GB	18.7 GB	20.9 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 256K context · ~203 tok/s
18.1 GB	18.2 GB	18.7 GB	24.5 GB	4-bit small — 4-bit small quantization	Fits: 256K context · ~203 tok/s
19.3 GB	19.4 GB	19.9 GB	25.3 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 256K context · ~197 tok/s
19.7 GB	19.8 GB	20.3 GB	25.6 GB	4-bit large — 4-bit large quantization	Fits: 256K context · ~196 tok/s
20.1 GB	20.2 GB	20.7 GB	28.8 GB	5-bit · legacy — 5-bit legacy quantization	Fits: 256K context · ~194 tok/s
22.1 GB	22.2 GB	22.6 GB	26.2 GB	5-bit small — 5-bit small quantization	Fits: 256K context · ~186 tok/s
22.9 GB	23 GB	23.4 GB	30.2 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Fits: 128K context · ~183 tok/s
23.3 GB	23.4 GB	23.8 GB	27.1 GB	5-bit large — 5-bit large quantization	Fits: 128K context · ~181 tok/s
26.4 GB	26.5 GB	27 GB	34.3 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 26.4 GB
31.9 GB	32 GB	32.5 GB	~31.6 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 32 GB
63.5 GB	63.6 GB	64.1 GB	~63.2 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 63.5 GB

Selected Q4_K_M

4-bit medium quantization

19.3GBat 16K context

Weights: 19.0 GB
KV cache: 0.1 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 4.7 GB free · ~197 tok/s

≥50 tok/s needs ≥130 GB/s · RTX 5090 ~244, RTX 3090 Ti ~197

Which Hardware Can Run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16?

Q4_K_M · 19.3 GB

NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 (Q4_K_M) requires 19.3 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 26+ GB is recommended. Using the full 262K context window can add up to 1.2 GB, bringing total usage to 20.5 GB. 8 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti. 122 devices with unified memory can run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16, including NVIDIA DGX Spark, NVIDIA Jetson AGX Thor Developer Kit, Mac Studio M1 Max (32 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060Doesn't fit — smallest quant needs 9.1 GB8 GB tier →

12 GB · RTX 3060 12GBIQ2_M · 11 GB · ~157 tok/s · 128K11~157128K12 GB tier →

16 GB · RTX 5060 Ti 16GBQ3_K_M · 15.8 GB · ~145 tok/s · 32K15.8~14532K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 19.3 GB · ~197 tok/s · 256K19.3~197256K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 19.3 GB · ~36 tok/s · 256K19.3~36256KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 19.3 GB · ~71 tok/s · 256K19.3~71256KMacs →

How good is NVIDIA Nemotron 3.5 Lightning 30B A3B BF16?

llmrun Score blends public benchmark boards into one number; 100 = the average open model.

Not enough data yet for an llmrun Score.

A Score needs results on enough independent boards. How the Composite Score is built

Benchmarks

LMArena Text
LMArena · Oct 2026Arena
1351.3
#76/187 open
Arena
LiveBench 2026 Coding
LiveBench · Sep 2026Independent
22.6
#15/15 open
Independent
LiveBench 2026 Long-Horizon
LiveBench · Sep 2026Independent
8.3
#15/15 open
Independent
LiveBench 2026 Design
LiveBench · Sep 2026Independent
43.9
#15/15 open
Independent
LiveBench 2026 Research
LiveBench · Sep 2026Independent
40.0
#15/15 open
Independent
LiveBench 2026 Reasoning
LiveBench · Sep 2026Independent
42.9
#15/15 open
Independent

All 8 NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 benchmarks

Alternatives to NVIDIA Nemotron 3.5 Lightning 30B A3B BF16

Nemotron family · 42 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Hy Memory 30B A3BTencent	—	19.1 GB	~184	same size
Llm Jp 4.1 32B A3b Thinkingllm-jp	—	20.1 GB	~188	same size
Llm Jp 4.1 33B Thinkingllm-jp	—	22.5 GB	~35	same size
Agnes 3.0 FlashAgnes-AI	—	22.1 GB	~35	same size
Muse Glimmer 30BMeta	—	19.6 GB	~36	same size
Hy MT2 30B A3BTencent	—	19.1 GB	~170	same size

Fine-tunes & variants(1)community builds of NVIDIA Nemotron 3.5 Lightning 30B A3B BF16

Nemotron 3.5 Lightning 30B A3B JANG 4M

Specifications

Publisher: NVIDIA
Family: Nemotron
Parameters: 31.6B
Active parameters: 2.4B per token
Architecture: NemotronHForCausalLM
Context Length: 262,144 tokens
Vocabulary Size: 131,072
Release Date: 2026-08-01
License: Other

Resources

Frequently Asked Questions

How much VRAM does NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 need?

NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 requires 19.3 GB of VRAM at Q4_K_M, or 63.5 GB at BF16. Full 262K context adds up to 1.2 GB (20.5 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 31.6B × 4.8 bits ÷ 8 = 18.9 GB

KV Cache + Overhead ≈ 0.4 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 1.6 GB (at full 262K context)

VRAM usage by quantization

Q4_K_M

19.3 GB

Q4_K_M + full context

20.5 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16?

Yes, at Q5_K_L (23.3 GB) or lower. Higher quantizations like Q6_K (26.4 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for NVIDIA Nemotron 3.5 Lightning 30B A3B BF16?

For NVIDIA Nemotron 3.5 Lightning 30B A3B BF16, Q4_K_M (19.3 GB) offers the best balance of quality and VRAM usage. Q4_K_L (19.7 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 9.1 GB.

VRAM requirement by quantization

IQ2_XXS

9.1 GB~53%

IQ3_S

13.8 GB~75%

IQ4_XS

17.3 GB~87%

Q4_K_M ★

19.3 GB~89%

Q5_0

20.1 GB~90%

BF16

63.5 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 on a Mac?

Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 19.3 GB of usable unified memory (RAM minus macOS overhead), from Mac Mini M5 Pro (24 GB).

Can I run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 locally?

Yes — NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is NVIDIA Nemotron 3.5 Lightning 30B A3B BF16?

At Q4_K_M, NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 can reach ~93 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~197 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 1.4 GB active ÷ (8000 × 0.65) = 0.278 ms, plus 52 layers × 0.055 ms = 2.860 ms, so 1000 ÷ 3.138 ms = ~319 tok/s

Estimated speed at Q4_K_M (19.3 GB)

NVIDIA B200

~319 tok/s

NVIDIA GeForce RTX 4090

~197 tok/s

NVIDIA B300

~319 tok/s

NVIDIA GH200 Grace Hopper Superchip

~302 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of NVIDIA Nemotron 3.5 Lightning 30B A3B BF16?

At Q4_K_M, the download is about 18.95 GB. The full-precision BF16 version is 63.16 GB. The smallest option (IQ2_XXS) is 8.68 GB.

Which GPUs can run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16?

8 consumer GPUs can run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 at Q4_K_M. Top options include NVIDIA GeForce RTX 5090, AMD Radeon RX 7900 XT, AMD Radeon RX 7900 XTX. 1 GPU has plenty of headroom for comfortable inference.

Which devices can run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16?

122 devices with unified memory can run NVIDIA Nemotron 3.5 Lightning 30B A3B BF16 at Q4_K_M (19.3 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.