NVIDIANemotronNemotronHPuzzleForCausalLM

NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 — Hardware Requirements & GPU Compatibility

ChatFunctions

NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16 is a deployment-optimized compression of NVIDIA's Nemotron-3-Super-120B-A12B, produced with Iterative Puzzle, a neural-architecture-search-based post-training compression framework, cutting the model from 120.7B total / 12.8B active parameters down to 75.3B total / roughly 9.3B active. It keeps the parent's hybrid architecture of interleaved Mamba, mixture-of-experts, and attention layers with multi-token prediction for faster generation, and the card reports about 2x higher server throughput on a single 8-GPU B200 node while only slightly trailing Nemotron-3-Super on reasoning, coding, agentic, and long-context benchmarks. It is a general-purpose reasoning and chat model intended for agents, chatbots, and RAG systems across seven languages, and needs a multi-GPU server to run. Context length defaults to 262,144 tokens in the published Hugging Face configuration, though the card states the architecture supports up to 1,048,576 tokens at higher memory cost. It is released under the OpenMDW License, version 1.1, a custom license for open model weights, and the card states it was published on July 6, 2026.

75.4B params · MoE9B active262K contextOtherJun 2026952 downloads71 likes2.7K quant downloads

Based on NVIDIA Nemotron 3 Super 120B A12B BF16

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.

NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 runs on a 64 GB+ Mac or a workstation GPU

Best local pick → Mac Pro M2 Ultra (64 GB) · Q4_K_M · 54.5 GB · ~50 tok/s

Macs from 48 GB (Q3_K_S · ~51 tok/s)

Other

Not set up for it?

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16

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

Recommended

Ollama

one command · needs Ollama installed

$ollama run hf.co/RemySkye/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-GGUF:Q4_K_M

Q4_K_M needs offload on the RTX 4090 · 30.5 GB short · 16K context

Download: NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-Q4_K_M.gguf51.60 GBRemySkye

install, OS, other quants

Other ways

each opens step-by-step

Run in cloud

Fits on RTX PRO 6000 (96 GB) (41 GB headroom) · Q4_K_M

Generation speed: ~129 tok/s; generation speed
Cost per 1M output tokens: $2.88; per 1M output tokens

Deployon Vast.aiBest optionfrom $1.34/hr

Compare GPUs →

or

How Much VRAM Does NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 Need?

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

Quantization	VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
	35.2 GB	45.5 GB	78.5 GB	31.5 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Does not fit: needs 40 GB
	36.3 GB	46.6 GB	79.5 GB	35.7 GB	3-bit small — 3-bit small quantization	Does not fit: needs 41.1 GB
	40.4 GB	50.7 GB	83.7 GB	40.6 GB	3-bit medium — 3-bit medium quantization	Does not fit: needs 45.2 GB
	41.4 GB	51.7 GB	84.7 GB	44.5 GB	4-bit · legacy — 4-bit legacy quantization	Does not fit: needs 46.3 GB
	42.5 GB	52.8 GB	85.8 GB	44.2 GB	3-bit large — 3-bit large quantization	Does not fit: needs 47.3 GB
	44.5 GB	54.9 GB	87.8 GB	43 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Does not fit: needs 49.4 GB
	46.6 GB	56.9 GB	89.9 GB	49.3 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Does not fit: needs 51.4 GB
	46.6 GB	56.9 GB	89.9 GB	46.3 GB	4-bit small — 4-bit small quantization	Does not fit: needs 51.4 GB
	49.7 GB	60 GB	93 GB	51.6 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Does not fit: needs 54.5 GB
	51.8 GB	62.1 GB	95.1 GB	54.2 GB	5-bit · legacy — 5-bit legacy quantization	Does not fit: needs 56.6 GB
	57.0 GB	67.3 GB	100.3 GB	54.9 GB	5-bit small — 5-bit small quantization	Does not fit: needs 61.8 GB
	57.0 GB	67.3 GB	100.3 GB	59.1 GB	5-bit · legacy+ — 5-bit legacy quantization with offset	Does not fit: needs 61.8 GB
	59.1 GB	69.4 GB	102.3 GB	58.7 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 63.9 GB
	68.4 GB	78.7 GB	111.7 GB	67.2 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 73.2 GB
	82.9 GB	93.2 GB	126.2 GB	83.5 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 87.7 GB
	165.8 GB	176.1 GB	209.1 GB	156.8 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 170.6 GB

Selected Q4_K_M

4-bit medium quantization

54.5GBat 16K context

Weights: 48.7 GB
KV cache: 5.5 GB
Overhead: 0.3 GB

Doesn't fit RTX 4090 (reference) · needs 30.5 GB more

≥50 tok/s needs ≥526 GB/s · M3 Ultra ~50

Or rent RTX PRO 6000 from $1.34/hr.

Which Hardware Can Run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16?

Q4_K_M · 49.7 GB

NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 (Q4_K_M) requires 49.7 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 65+ GB is recommended. No consumer GPU has enough memory. 52 devices with unified memory can run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Pro M2 Ultra (64 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8–16 GB · RTX 4060, RTX 3060 12GB, RTX 5060 Ti 16GBDoesn't fit — smallest quant needs 40 GB8 GB tier →12 GB tier →16 GB tier →

24 GB · RTX 4090referenceDoesn't fit — smallest quant needs 40 GB24 GB tier →

Mac 32 GB · Mac Mini M4Doesn't fit — smallest quant needs 40 GBMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 54.5 GB · ~40 tok/s · 32K54.5~4032KMacs →

Alternatives to NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16

Nemotron family · 42 models →

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Llama 3.3 70B InstructMeta	79.7	51.3 GB	—needs > 24 GB	same size
Kolibri 1 BF16Aleph-Alpha	—	47.9 GB	—needs > 24 GB	same size
AliceAI Foundation 80B A3B Baseyandex	—	49.9 GB	—needs > 24 GB	same size
Qwen3 Next 80B A3B InstructAlibaba	—	22.9 GB	~200	same size
Qwen2.5 VL 72B InstructAlibaba	—	49.7 GB	—needs > 24 GB	same size
Qwen2 VL 72B InstructAlibaba	—	49.7 GB	—needs > 24 GB	same size

Specifications

Publisher: NVIDIA
Family: Nemotron
Parameters: 75.4B
Active parameters: 9B per token
Architecture: NemotronHPuzzleForCausalLM
Context Length: 262,144 tokens
Vocabulary Size: 131,072
Release Date: 2026-06-24
License: Other

Resources

Frequently Asked Questions

How much VRAM does NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 need?

NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 requires 49.7 GB of VRAM at Q4_K_M, or 165.8 GB at BF16.

VRAM = Weights + KV Cache + Overhead

Weights = 75.4B × 4.8 bits ÷ 8 = 45.2 GB

KV Cache + Overhead ≈ 4.5 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.

VRAM usage by quantization

Q4_K_M

49.7 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 5090 run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16?

No — NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 requires at least 35.2 GB at Q2_K, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.

What's the best quantization for NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16?

For NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16, Q4_K_M (49.7 GB) offers the best balance of quality and VRAM usage. Q5_0 (51.8 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 35.2 GB.

VRAM requirement by quantization

Q2_K

35.2 GB~75%

Q3_K_L

42.5 GB~86%

Q4_K_M ★

49.7 GB~89%

Q5_0

51.8 GB~90%

Q5_1

57.0 GB~92%

BF16

165.8 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 on a Mac?

Yes — Mac Studio M5 Max (48 GB) and 55 other Macs can run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16, the smallest at Q3_K_S. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 49.7 GB of usable unified memory (RAM minus macOS overhead), from Mac Pro M2 Ultra (64 GB).

Can I run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 locally?

Yes — NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16?

At Q4_K_M, NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 can reach ~92 tok/s on AMD Instinct MI350X. 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 → 5.9 GB active ÷ (8000 × 0.65) = 1.142 ms, plus 48 layers × 0.055 ms = 2.640 ms, so 1000 ÷ 3.782 ms = ~264 tok/s

Estimated speed at Q4_K_M (49.7 GB)

NVIDIA B200

~264 tok/s

NVIDIA B300

~264 tok/s

NVIDIA GH200 Grace Hopper Superchip

~222 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 Labs 3 Puzzle 75B A9B BF16?

At Q4_K_M, the download is about 45.21 GB. The full-precision BF16 version is 150.71 GB. The smallest option (Q2_K) is 32.03 GB.

Which GPUs can run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16?

No single consumer GPU has enough VRAM to run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 at Q4_K_M (49.7 GB). Multi-GPU or professional hardware is required.

Which devices can run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16?

52 devices with unified memory can run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16 at Q4_K_M (49.7 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (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.

Run NVIDIA Nemotron Labs 3 Puzzle 75B A9B BF16

54.5 GB VRAM at 16K context · needs offload on the RTX 4090 (30.5 GB short)

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/RemySkye/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-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/RemySkye/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/RemySkye/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-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 RemySkye/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-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/RemySkye/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-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 RemySkye/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-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-Labs-3-Puzzle-75B-A9B-BF16 --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~166 GB of GPU memory.

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.
$RemySkye/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-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
RTX PRO 6000 Max-Q (96 GB)CheapestCheapest41 GB free	~129 tok/s	$2.88	Deploy on Vast.aifrom $1.34/hr (opens in new tab)	96 GB · 41 GB free · ~129 tok/s · $2.88 / 1M tokens
A100 80GB25 GB free	~140 tok/s	$2.75	Deploy on RunPodfrom $1.39/hr (opens in new tab)	25 GB free · ~140 tok/s · $2.75 / 1M tokens
RTX PRO 6000 (96 GB)41 GB free	~129 tok/s	$3.02	Deploy on Vast.aifrom $1.41/hr (opens in new tab)	96 GB · 41 GB free · ~129 tok/s · $3.02 / 1M tokens
A100 80GB25 GB free	~140 tok/s	$3.01	Deploy on Vast.aifrom $1.52/hr (opens in new tab)	25 GB free · ~140 tok/s · $3.01 / 1M tokens
RTX PRO 6000 (96 GB)41 GB free	~129 tok/s	$3.63	Deploy on RunPodfrom $1.69/hr (opens in new tab)	96 GB · 41 GB free · ~129 tok/s · $3.63 / 1M tokens
RTX PRO 6000 Server (96 GB)41 GB free	~120 tok/s	$3.93	Deploy on RunPodfrom $1.69/hr (opens in new tab)	96 GB · 41 GB free · ~120 tok/s · $3.93 / 1M tokens
H100 (80 GB)25 GB free	~186 tok/s	$4.01	Deploy on RunPodfrom $2.69/hr (opens in new tab)	80 GB · 25 GB free · ~186 tok/s · $4.01 / 1M tokens
H200 (141 GB)86 GB free	~220 tok/s	$5.30	Deploy on Vast.aifrom $4.20/hr (opens in new tab)	141 GB · 86 GB free · ~220 tok/s · $5.30 / 1M tokens

Compare GPUs →