Llama 3.2 1B Instruct

Llama 3.2 1B Instruct is the smallest instruction-tuned text model in Meta's Llama 3.2 release. Its Hugging Face model card lists about 1.23B parameters, text in and text out, and a release date of September 25, 2024. It is aimed at assistant-style chat and light agentic tasks on limited compute.

Key Features

• Context length: 128k tokens for the standard model. The quantized variant has an 8k context.
• Languages: officially supports English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai. Developers may fine-tune it for other languages under the license and acceptable use policy.
• Training data: up to 9T pretraining tokens from public online sources, with a knowledge cutoff of December 2023.
• Architecture: uses Grouped-Query Attention (GQA) and shared embeddings. Post-training uses supervised fine-tuning and preference-based alignment.
• On-device options: quantized versions use 4-bit groupwise weights (group size 32) with 8-bit dynamic activations, and the card also describes quantization-aware training with LoRA and SpinQuant variants.

Intended Use

Meta lists assistant-style chat, agentic applications such as knowledge retrieval and summarization, writing assistance, and query rewriting. The quantized versions target on-device use with limited compute.

License

Llama 3.2 Community License, a custom commercial license that includes an acceptable use policy. Check the license terms on the model card before shipping a product.

Use Cases

• Running a small assistant locally or on a phone-class device.
• Query rewriting and summarization in a retrieval pipeline.
• Fast prototypes and fine-tuning experiments where a 1B model is enough.

Limitation: a 1B model is weaker than larger Llama models on complex reasoning, and the official language list is only eight languages. Test it on your own tasks before choosing it over a bigger model.

Getting Started

1. Request access on the Hugging Face model card and accept the license.
2. Load meta-llama/Llama-3.2-1B-Instruct with Hugging Face transformers, or use a quantized variant for a device.
3. Use the chat template, and measure quality on your own prompts.

FAQ

How many parameters does Llama 3.2 1B Instruct have?

About 1.23B, per the model card.

What is its context length?

128k tokens for the standard model and 8k for the quantized variant.

Which languages are officially supported?

English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai.

Can I use it commercially?

Under the Llama 3.2 Community License, which includes an acceptable use policy. Read the terms on the card.

Alternatives

• Llama 3.1 8B Instruct: a larger Llama for stronger quality.
• Llama 3.2 Vision: the vision-capable Llama 3.2 models.
• Llama Guard 3: Meta's safety classifier for moderating Llama outputs.

Conclusion

Llama 3.2 1B Instruct is a compact open-weights model for on-device and low-cost text tasks. Start from the model card and compare other options in the models category.