---
license: apache-2.0
language:
- en
tags:
- security
- cybersecurity
- mlx
- mlx-bf16
base_model: viettelsecurity-ai/security-llama3.2-3b
base_model_relation: quantized
pipeline_tag: text-generation
library_name: mlx
quantized_by: ahmedandaloes
---

# security-llama3.2-3b — MLX bf16

Full-precision (bf16) [MLX](https://github.com/ml-explore/mlx) build of
[**viettelsecurity-ai/security-llama3.2-3b**](https://huggingface.co/viettelsecurity-ai/security-llama3.2-3b), for fast local inference on Apple Silicon.

- **Precision:** bf16 — no quantization, identical weights to source.
- Weights unchanged from source — format + precision conversion only.
- Converted with `mlx-lm`.

## Builds
| Build |
|-------|
| [MLX-4bit](https://huggingface.co/ahmedandaloes/security-llama3.2-3b-MLX-4bit) |
| [MLX-6bit](https://huggingface.co/ahmedandaloes/security-llama3.2-3b-MLX-6bit) |
| [MLX-8bit](https://huggingface.co/ahmedandaloes/security-llama3.2-3b-MLX-8bit) |
| [MLX-bf16](https://huggingface.co/ahmedandaloes/security-llama3.2-3b-MLX-bf16) |

GGUF builds: [prithivMLmods/Security-Llama3.2-3B-GGUF](https://huggingface.co/prithivMLmods/Security-Llama3.2-3B-GGUF).

## Usage
```bash
pip install mlx-lm
```
```python
from mlx_lm import load, generate
model, tok = load("ahmedandaloes/security-llama3.2-3b-MLX-bf16")
p = tok.apply_chat_template([{"role":"user","content":"Name a common web vulnerability."}], add_generation_prompt=True)
print(generate(model, tok, prompt=p, max_tokens=200, verbose=True))
```

## Attribution
Source: [viettelsecurity-ai/security-llama3.2-3b](https://huggingface.co/viettelsecurity-ai/security-llama3.2-3b). License per source (Apache-2.0 assumed; verify).
MLX build for the Apple Silicon community. For **authorized** security work only.