Llama-3.2-3B-Instruct-MNN

Pre-converted Llama 3.2 3B Instruct in MNN format for on-device inference with TokForge.

Original model by Meta — converted to MNN Q4 for mobile deployment.

Model Details

Architecture	Llama 3.2 (standard attention, 28 layers, GQA 24Q/8KV)
Parameters	3B (4-bit quantized)
Format	MNN (Alibaba Mobile Neural Network)
Quantization	W4A16 (4-bit weights, block size 128)
Vocab	128,256 tokens
Source	meta-llama/Llama-3.2-3B-Instruct

Description

Meta's official Llama 3.2 3B Instruct — the compact powerhouse of the Llama family. Designed specifically for edge and mobile deployment. Excellent instruction following in a package that runs on 8GB+ phones. Supports 128K context and 8 languages.

Files

File	Description
llm.mnn	Model computation graph
llm.mnn.weight	Quantized weight data (Q4, block=128)
llm_config.json	Model config with Jinja chat template
tokenizer.txt	Tokenizer vocabulary
config.json	MNN runtime config

Usage with TokForge

This model is optimized for TokForge — a free Android app for private, on-device LLM inference.

1. Download TokForge from the Play Store
2. Open the app → Models → Download this model
3. Start chatting — runs 100% locally, no internet required

Recommended Settings

Setting	Value
Backend	OpenCL (Qualcomm) / Vulkan (MediaTek) / CPU (fallback)
Precision	Low
Threads	4
Thinking	Off (or On for thinking-capable models)

Performance

Actual speed varies by device, thermal state, and generation length. Typical ranges for this model size:

Device	SoC	Backend	tok/s
RedMagic 11 Pro	SM8850	OpenCL	~25-28 tok/s
Lenovo TB520FU	SM8650	OpenCL	~18-22 tok/s

Attribution

This is an MNN conversion of Llama 3.2 3B Instruct by Meta. All credit for the model architecture, training, and fine-tuning goes to the original author(s). This conversion only changes the runtime format for mobile deployment.

Limitations

• Intended for TokForge / MNN on-device inference on Android
• This is a runtime bundle, not a standard Transformers training checkpoint
• Quantization (Q4) may slightly reduce quality compared to the full-precision original
• Abliterated/uncensored models have had safety filters removed — use responsibly

Community

• Website: tokforge.ai
• Discord: Join our Discord
• GitHub: TokForge on GitHub

Export Details

Converted using MNN's llmexport pipeline:

python llmexport.py --path meta-llama/Llama-3.2-3B-Instruct --export mnn --quant_bit 4 --quant_block 128