Llama 3 at a glance
Llama 3 used about seven times more training data than Llama 2, a new 128K-token vocabulary and wide cloud availability, and it powered the Meta AI assistant.
What changed with Llama 3
Meta released Llama 3 on April 18, 2024, with 8B and 70B parameter models, each in pretrained and instruction-tuned versions. Meta said they were trained on more than 15 trillion tokens, about seven times the Llama 2 dataset, using sequences of 8,192 tokens and a new tokenizer with a 128K vocabulary that encodes text with up to 15% fewer tokens.
The models launched across AWS, Google Cloud, Microsoft Azure, Databricks, Hugging Face, Kaggle, IBM watsonx, NVIDIA NIM and Snowflake. Meta AI, the assistant in Facebook, Instagram, WhatsApp and Messenger, was built on Llama 3, and Meta said models over 400B parameters were still in training.
Where to compare Llama 3
Independent leaderboards and official resources for comparing this model yourself. Rankings change often; each site shows its own current data.
General leaderboards
Why Llama 3 matters for businesses
For businesses, Llama 3 70B brought open-weight quality close to the paid models of the time, and the 8B version was strong enough for many self-hosted tasks. Its availability on every major cloud made it easy to try without new contracts.
Sources for this story
- Introducing Meta Llama 3: The most capable openly available LLM to date (opens in a new tab)Official
Meta · ai.meta.com
- Llama 3 70B model page (opens in a new tab)Benchmark
Artificial Analysis · artificialanalysis.ai
Last verified: . IA Compa summarizes public announcements for business readers. We are not affiliated with the companies mentioned. Always confirm details, pricing, and availability with the official source.
- #llm
- #meta
- #llama
- #open-weights