README.md

---
license: llama3.2
base_model: meta-llama/Llama-3.2-3B-Instruct
tags:
  - rkllm
  - rk3588
  - rockchip
  - npu
  - quantized
  - llama
language:
  - en
---

# Llama-3.2-3B-Instruct — RKLLM v1.2.3 (w8a8, RK3588)

RKLLM conversion of [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) for Rockchip RK3588 NPU inference.

Converted with RKLLM Toolkit v1.2.3. This is a standard instruct model — it does **not** produce `<think>` reasoning blocks.

## Key Details

| Property | Value |
|----------|-------|
| Base Model | meta-llama/Llama-3.2-3B-Instruct |
| Toolkit Version | RKLLM Toolkit v1.2.3 |
| Runtime Version | RKLLM Runtime ≥ v1.2.1 (v1.2.3 recommended) |
| Quantization | w8a8 (8-bit weights, 8-bit activations) |
| Quantization Algorithm | normal |
| Target Platform | RK3588 |
| NPU Cores | 3 |
| Max Context Length | 8192 tokens |
| Optimization Level | 0 |
| Thinking Mode | ❌ Not supported |
| Languages | English (+ multilingual inherited from Llama 3.2) |

## Why This Model?

Llama 3.2 3B Instruct is Meta's latest compact instruction model. It brings a different architecture and training lineage to the RK3588 NPU lineup — strong at instruction following, coding, and general reasoning without the overhead of a thinking/reasoning mode.

At ~3B parameters it sits between the 1.7B and 4B Qwen3 models, offering a useful middle ground.

## Hardware Tested

- **Orange Pi 5 Plus** — RK3588, 16GB RAM, Armbian Linux
- RKNPU driver 0.9.8
- RKLLM Runtime v1.2.3

## Usage

### With the official RKLLM API demo

```bash
# Clone the runtime
git clone https://github.com/airockchip/rknn-llm.git
cd rknn-llm/examples/rkllm_api_demo

# Run (aarch64)
./build/rkllm_api_demo /path/to/Llama-3.2-3B-Instruct-rk3588-w8a8.rkllm 4096 8192
```

### With the RKLLM API Server

Download and place in `~/models/`:

```bash
mkdir -p ~/models/Llama-3.2-3B-Instruct
cd ~/models/Llama-3.2-3B-Instruct
git lfs install && git clone https://huggingface.co/GatekeeperZA/Llama-3.2-3B-Instruct-RKLLM-v1.2.3 .
```

The server auto-detects the model as `instruct` capability (no thinking). Use with [GatekeeperZA/RKLLM-API-Server](https://github.com/GatekeeperZA/RKLLM-API-Server).

## Conversion Script

```python
from rkllm.api import RKLLM

llm = RKLLM()
llm.load_huggingface(model="meta-llama/Llama-3.2-3B-Instruct", device="cpu", dtype="float32")
llm.build(
    do_quantization=True,
    optimization_level=0,
    quantized_dtype="w8a8",
    quantized_algorithm="normal",
    target_platform="rk3588",
    num_npu_core=3,
    max_context=8192,
)
llm.export_rkllm("./Llama-3.2-3B-Instruct-rk3588-w8a8.rkllm")
```

> **WSL2 note:** Requires ≥16GB WSL2 memory (`memory=16GB` in `~/.wslconfig`). Write the output to a native Linux path (`/home/user/`) first, then copy to `/mnt/` — writing directly to the Windows mount triggers OOM during the export phase.

## File Listing

| File | Description |
|------|-------------|
| `Llama-3.2-3B-Instruct-rk3588-w8a8.rkllm` | Quantized model for RK3588 NPU |

## Compatibility Notes

- Minimum runtime: RKLLM Runtime v1.2.1. v1.2.3 recommended.
- RKNPU driver: ≥ 0.9.6
- SoCs: RK3588 / RK3588S (3 NPU cores). Not compatible with RK3576 (2 cores) without reconversion.
- RAM: ~3.5GB loaded. Runs comfortably on 8GB+ boards.

## Acknowledgements

- Meta / FAIR for the Llama 3.2 base model
- Rockchip / airockchip for the RKLLM toolkit and runtime
- Converted by [GatekeeperZA](https://huggingface.co/GatekeeperZA)