What is Llama 3?
Llama 3 — Meta's open-source flagship model
Llama 3 is an open-source large language model developed by Meta. Released in 2024, it quickly became the most popular model in the Ollama ecosystem due to its strong general-purpose performance and reasonable hardware requirements. The 8B parameter variant delivers quality comparable to much larger models on conversational and reasoning tasks.
On Windows, Llama 3 runs via Ollama with NVIDIA CUDA or AMD DirectML GPU acceleration. The 8B model fits comfortably in 6–8 GB VRAM and runs at 40–80 tokens per second on a mid-range GPU.
Quick start
Run Llama 3 on Windows
# Pull the default 8B model:
C:\> ollama pull llama3
pulling manifest...
pulling 8934d96d3f08... 100% ████████ 4.7 GB
success
# Start an interactive session:
C:\> ollama run llama3
>>> Hello! Can you help me write a Python script?
Of course! What would you like the script to do?
Variants
Llama 3 variants and hardware requirements
| Variant | Parameters | Size | Min VRAM | Recommended for |
|---|---|---|---|---|
llama3 | 8B | 4.7 GB | 6 GB | Most users — best starting point |
llama3:8b-instruct-q4_K_M | 8B Q4 | 4.4 GB | 5 GB | Lower VRAM, slightly faster |
llama3:8b-instruct-q8_0 | 8B Q8 | 8.5 GB | 10 GB | Higher quality, needs more VRAM |
llama3:70b | 70B | 40 GB | 48 GB | High-end workstations only |
For most Windows users,
ollama pull llama3 (the default 8B Q4 model) is the right choice. It runs on any GPU with 6+ GB VRAM and on CPU-only systems.Use cases
What Llama 3 is good at
Llama 3 8B performs well across a broad range of tasks:
- Writing and editing — drafting emails, documents, summaries
- Coding — writing, explaining and debugging code in most languages
- Q&A and reasoning — answering questions based on provided context
- Summarisation — condensing long documents into key points
- General conversation — the model is instruction-tuned for interactive use
FAQ
Llama 3 questions
What is the difference between Llama 3 and Llama 3.1?
Llama 3.1 (released July 2024) extends context length to 128K tokens and adds improved multilingual support. Pull it with
ollama pull llama3.1. For most tasks the difference is small; Llama 3 8B is faster and uses less VRAM.Is Llama 3 censored?
The default Llama 3 instruct model has safety guardrails trained by Meta. There are uncensored variants in the Ollama library if you need them for specific research use cases. Ollama itself imposes no additional restrictions.
Can I run Llama 3 on CPU only?
Yes. CPU-only speed is 5–12 tokens/s on a modern multi-core CPU, which is usable for short tasks. For longer conversations a GPU is strongly recommended.
How is Llama 3 licensed?
Llama 3 uses Meta's custom Llama 3 Community License. It is free for research and commercial use for products with fewer than 700 million monthly active users. Check the full license at llama.meta.com.