FastFlowLM/Llama-3.2-3B-NPU2

🦙 LLaMA 3.2 (3B) – Optimized for FastFlowLM on AMD Ryzen™ AI NPU (XDNA2 Only)

Model Summary

This model is a variant of Meta AI’s LLaMA 3.2 3B Instruct release. It preserves the original architecture and weights, with potential optimizations via quantization, low-level tuning, or runtime enhancements tailored for NPUs using FastFlowLM.

⚠️ This model is subject to Meta’s LLaMA 3 license. You must accept Meta’s terms to use or download it.

📝 License & Usage Terms

Meta LLaMA 3 License

• —Governed by Meta AI's LLaMA 3 license: 👉 https://ai.meta.com/llama/license/

• —Key restrictions include:
• —No commercial use without express permission from Meta
• —Redistribution must follow Meta’s guidelines
• —Attribution to Meta is required

Redistribution Notice

• —This repository does not contain Meta’s original weights.
• —You must obtain the base weights directly from Meta: 👉 https://huggingface.co/meta-llama

If Fine-tuned

If this version includes any fine-tuning or post-training modification:

• —Base Model License: Meta’s LLaMA 3 License
• —Derivative Weights License: [e.g., CC-BY-NC-4.0, MIT, custom]
• —Training Dataset License(s):
• —[Dataset A] – [license]
• —[Dataset B] – [license]

Users are responsible for verifying the legality of dataset use and redistribution.

Intended Use

• —Target Applications: On-device experimentation, local LLM inference, academic research
• —Exclusions: Do not use in commercial products, production systems, or critical tasks without proper evaluation and license compliance

Limitations & Risks

• —May hallucinate or output biased content
• —Knowledge is frozen as of the base model's training cutoff
• —Not evaluated for high-stakes or real-time applications

Citation

@misc{touvron2024llama3,
  title={LLaMA 3: Open Foundation and Instruction Models},
  author={Touvron, Hugo and others},
  year={2024},
  url={https://ai.meta.com/llama/}