Uniform MLX-native affine int4 (group size 32): every quantized tensor — routed MoE experts, Mamba in/out projections, attention q/k/v/o, shared experts, MoE latent projections — is 4-bit; router gate, conv1d, embeddings and lm_head stay bf16.

It trades fidelity for speed versus mlx-community/Nemotron-3-Ultra-550B-A55B (which keeps the mixing path at int8): ~32% faster decode, at an output-logit cosine of 0.9906 vs the int8-mixing model's 0.9958 (top-1 token agreement 97.9% vs 98.7%) against the source model.

Use with mlx

pip install mlx-lm

python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/Nemotron-3-Ultra-550B-A55B-4bit")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

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