Meta Llama 3
Meta Llama 3 is an AI tool that allows users to build sophisticated AI technologies. It comes with the option of 8B and 70B pretrained and ...
Last verified:
What is Meta Llama 3?
Llama 3 represents a family of state-of-the-art open-source large language models developed by Meta, available in sizes from 8B to 70B parameters, including both pre-trained and instruction-tuned variants. These models excel in a wide range of tasks such as general knowledge, multilingual understanding, coding, math, reasoning, tool use, and long-context processing, outperforming many competing open models on over 150 benchmarks. They support up to 8192 token context lengths and are designed for easy local inference, fine-tuning, distillation, and deployment anywhere, empowering individuals, researchers, creators, and businesses to innovate responsibly.
Key features include superior performance in MMLU (86.0+), HumanEval coding (80.5+), MATH (67.8+), and multilingual tasks, with instruction-tuned versions optimized for dialogue and chat applications using a specific chat format with special tokens like <|begin_of_text|>, <|start_header_id|>, and <|eot_id|>. The models offer the cheapest API pricing among providers at $0.1 per 1M input tokens and $0.4 per 1M output tokens. Extensive safety evaluations, human assessments, and resources like the Responsible Use Guide ensure ethical deployment.
Llama 3 is ideal for developers building smarter applications, researchers advancing AI, and enterprises scaling ideas with flexible, high-performing open models. Model weights and code are downloadable after license acceptance from the Meta Llama site or Hugging Face, with quick-start inference scripts for PyTorch/CUDA environments.
Meta Llama 3 pricing
Pricing model: Free
Model weights are free to download after license acceptance; no direct paid plans on site. API access available through providers with Llama 3 at $0.1 per 1M input tokens and $0.4 per 1M output tokens, noted as cheapest among providers based on Artificial Analysis data.
Meta Llama 3 pros
- State-of-the-art performance on 150+ benchmarks
- Open-source weights freely downloadable
- Multiple sizes: 8B and 70B parameters
- Pre-trained and instruction-tuned variants
- Supports 8192 token context length
- Excels in multilingual tasks like MGSM
- Top coding scores on HumanEval 80.5%
- Strong math reasoning on MATH 67.8%
- High instruction following IFEval 87.5%
- Cheapest API at $0.1/M input tokens
- Optimized for dialogue with chat format
- Easy local inference with torchrun
- Fine-tunable and distillable anywhere
- Outperforms many open-source chat models
- Backed by extensive human evaluations
- Rapid growth with millions of downloads
- Hugging Face integration for transformers
Meta Llama 3 cons
- Requires license acceptance for downloads
- Download links expire after 24 hours
- High hardware needs for 70B model
- Model-parallel setup for larger sizes
- No use for improving other LLMs
- Litigation against Meta terminates license
- Must display 'Built with Meta Llama 3'
- Safety risks require Responsible Use Guide
- Deprecated GitHub repo for new versions
Frequently asked questions about Meta Llama 3
How do I download Llama 3 models?
Visit the Meta Llama website, accept the license to receive a signed URL via email, then run the download.sh script with wget and md5sum installed. Links expire after 24 hours; re-request if needed. Alternatively, access on Hugging Face after approval for transformers or native formats.
What are the model sizes available?
Llama 3 includes 8B and 70B parameter models in pre-trained and instruction-tuned versions. Use 8B with nproc_per_node 1, 70B with 8 for model-parallel inference.
What is the context length supported?
All Llama 3 models support up to 8192 tokens. Adjust max_seq_len and max_batch_size based on hardware when running inference.
How to run instruction-tuned models?
Use specific chat format starting with <|begin_of_text|>, followed by messages with <|start_header_id||>role<|end_header_id|> content <|eot_id|>. Run example_chat_completion.py with torchrun and appropriate ckpt_dir.
What license applies to Llama 3?
Meta Llama 3 Community License allows use for research and commercial purposes. Requires attributing 'Built with Meta Llama 3', retaining notice in copies, and prohibits using outputs to improve other LLMs.
What benchmarks does Llama 3 excel in?
Leads in MMLU Chat 86.0+, MMLU PRO 68.9+, IFEval 92.1+, HumanEval 88.4+, Multilingual MGSM 91.1+, with competitive scores vs GPT-4o and Claude 3.5 Sonnet across general, code, math, reasoning.
Can I use Llama 3 commercially?
Yes, licensed for commercial entities. Distribute with agreement copy, display attribution, include 'Llama 3' in derivative model names, and follow acceptable use policy.
What hardware is needed for inference?
PyTorch/CUDA environment required. 8B runs on single GPU with nproc_per_node 1; 70B needs model-parallel with 8. Pre-allocate cache via max_seq_len and max_batch_size.
How does Llama 3 handle safety?
Extensive evaluations on 150+ benchmarks and human assessments. Use Responsible Use Guide, add safety classifiers as in llama-cookbook. Report risky outputs to developers.facebook.com/llama_output_feedback.
Where to find more examples and code?
Check llama-cookbook for community scripts, integrations, safety checkers. Main repo deprecated; use llama-models, PurpleLlama, llama-toolchain for latest Llama Stack components.