Instructions to use georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit with MLX:

# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Notebooks
Google Colab
Kaggle
Local Apps Settings
LM Studio

How to use georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit"

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit"

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit

Run Hermes

hermes

OpenClaw new

How to use georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit"

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

MLX LM

How to use georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit with MLX LM:

Generate or start a chat session

# Install MLX LM
uv tool install mlx-lm
# Interactive chat REPL
mlx_lm.chat --model "georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit"

Run an OpenAI-compatible server

# Install MLX LM
uv tool install mlx-lm
# Start the server
mlx_lm.server --model "georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit"
# Calling the OpenAI-compatible server with curl
curl -X POST "http://localhost:8000/v1/chat/completions" \
   -H "Content-Type: application/json" \
   --data '{
     "model": "georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit",
     "messages": [
       {"role": "user", "content": "Hello"}
     ]
   }'

Nemotron-Labs-3-Puzzle-75B-A9B — MLX 6-bit

MLX 6-bit affine quantization of nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16, converted for Apple Silicon.

Runtime setup

Stock mlx-lm (0.31.x) can't load this model yet — it crashes with uniform(): incompatible function arguments because NVIDIA's Puzzle architecture uses different MoE dims per layer, and mlx-lm assumes they're all the same.

About the model

Puzzle-75B-A9B is NVIDIA's deployment-optimized compression of Nemotron-3-Super-120B-A12B. It's a hybrid Mamba-2 / Attention / LatentMoE architecture (88 backbone layers) with heterogeneous per-layer expert configs - MoE intermediate sizes vary 1280–2688 and active experts per token vary 4–22 across layers. 75.3B total / 9.3B active parameters.

This conversion

Source: nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16
Format: MLX affine 6-bit, group size 64 (~6.5 bpw)
Size on disk: ~57 GB
Converted with: mlx-lm 0.31.2 + mlx 0.31.1 (CUDA backend on a Blackwell), plus local patches to nemotron_h.py to support Puzzle's heterogeneous per-layer MoE dims, LatentMoE fc1_latent_proj / fc2_latent_proj, and the model. → backbone. prefix in NVIDIA's checkpoints.
MTP weights: mtp.safetensors (5.9 GB) from the source is included in this repo but is not currently used at inference time — mlx-lm has no Nemotron-H MTP path yet (tracking mlx-lm#1161). The tensors are preserved here so they'll be available whenever native speculative-decoding support lands.

License

Governed by the NVIDIA Open Model License. Derivative of NVIDIA's Nemotron-3 family. "Nemotron" is a trademark of NVIDIA Corporation. Not affiliated with or endorsed by NVIDIA.

Downloads last month: 979

Safetensors

Model size

75B params

Tensor type

BF16

U32

F32

MLX

Hardware compatibility

6-bit

Model tree for georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit

Base model

nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16

Finetuned

nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16

Quantized

(5)

this model

Collection including georgeis55/Nemotron-Labs-3-Puzzle-75B-A9B-MLX-6bit

Everyday MLX

Collection

A collection of MLX models optimized for Mac devices • 7 items • Updated 21 days ago