---
title: Nemotron-3-Ultra-550B-A55B-4bit
canonical_url: "https://www.modelscope.cn/models/mlx-community/Nemotron-3-Ultra-550B-A55B-4bit"
md_url: "https://www.modelscope.cn/models/mlx-community/Nemotron-3-Ultra-550B-A55B-4bit.md"
repository: mlx-community/Nemotron-3-Ultra-550B-A55B-4bit
last_updated: 2026-06-06
license: other
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - nemotron_h
architectures:
  - NemotronHForCausalLM
base_model:
  - nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
base_model_relation: quantized
parameters: 104.9B
tensor_type:
  - F32
  - BF16
  - U32
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 15
stars: 0
tags:
  - mlx
  - nemotron
---

# Nemotron-3-Ultra-550B-A55B-4bit

> Nemotron-3-Ultra-550B-A55B-4bit - mlx-community 在 ModelScope 开源的模型。mlx-community/Nemotron-3-Ultra-550B-A55B-4bit

mlx-community/Nemotron-3-Ultra-550B-A55B-4bit 是 ModelScope 魔搭社区上的 104.9B 参数text-generation模型，采用 other 许可，基于 nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 构建。

- **Repository**: mlx-community/Nemotron-3-Ultra-550B-A55B-4bit
- **License**: other
- **Tasks**: text-generation
- **Parameters**: 104.9B
- **Base model**: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
- **Tags**: mlx, nemotron
- **Downloads**: 15
- **Stars**: 0
- **Last updated**: 2026-06-06

Source: https://www.modelscope.cn/models/mlx-community/Nemotron-3-Ultra-550B-A55B-4bit

---

# mlx-community/Nemotron-3-Ultra-550B-A55B-4bit

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](https://huggingface.co/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

```bash
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)
```