Meta Llama 3

Available · Language · Milestone

Meta Llama 3 is the third generation of Meta's Llama language models, released on 18 April 2024 as open weights in 8B and 70B sizes, each pretrained and instruction-tuned, under the custom Llama 3 Community License. Meta also offered it through cloud partners and said its Meta AI assistant was built with Llama 3 technology. [1] [2] [3] Primary source

Timeline of Meta Llama 3 →

Claims and evidence

  • Released Primary source[1]page date; introduction [2]Model Release Date [4]README, Llama Models table
  • Status AvailableMeta publishes no deprecation page for Llama models; the weights of both sizes are still listed in Meta's Hugging Face organisation (gated, license acceptance required) as seen on 2026-10-01. Primary source[3]
  • Successor of Llama 2 Primary source[1]introduction; State-of-the-art performance
  • Change · Architecture Uses a new tokenizer with a 128K-token vocabulary, which Meta says encodes text with fewer tokens, and grouped query attention in both sizes.Compared with Llama 2 Primary source[1]Model architecture
  • Change · Context length Context length increased from 4K tokens for Llama 2 to 8K tokens.Compared with Llama 2 Primary source[4]README, Llama Models table
  • Change · Training data Meta says pretraining used over 15 trillion tokens, about seven times the Llama 2 dataset, with four times more code.Compared with Llama 2 Primary source[1]Training data
  • Change · Size Released in 8B and 70B sizes, where Llama 2 came in 7B, 13B and 70B.Compared with Llama 2 Primary source[4]README, Llama Models table
  • Input text Primary source[2]Model Details, Input
  • Output text, code Primary source[2]Model Details, Output
  • Open weights YesWeights under the custom Llama 3 Community License (a custom commercial license per the model card), after accepting the license. Primary source[1]Try Meta Llama 3 today [3]
  • Context window 8k tokens Primary source[2]Model Details table, Context length [1]Model architecture (sequences of 8,192 tokens)
  • Access Open-weights download, cloud partner, consumer appConsumer access through the Meta AI assistant, which Meta says is built with Llama 3 technology. Primary source[1]introduction; What's next for Llama 3?; Try Meta Llama 3 today

Lineage

Predecessors

Successors

Based on

Not derived from another model.

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

Variants

Meta Llama 3 8B

Same dates as Meta Llama 3.

  • Variant Inline variant in this record.Released as a pretrained model (Meta-Llama-3-8B) and an instruction-tuned model (Meta-Llama-3-8B-Instruct). Primary source[1] [2]Model Details [3]

Also known as: Meta-Llama-3-8B, Meta-Llama-3-8B-Instruct, Llama 3 8B

Differs in:

  • Parameters: 8B Evidence not assessed [2]Model Details table, Params

Meta Llama 3 70B

Same dates as Meta Llama 3.

  • Variant Inline variant in this record.Released as a pretrained model (Meta-Llama-3-70B) and an instruction-tuned model (Meta-Llama-3-70B-Instruct). Primary source[1] [2]Model Details [3]

Also known as: Meta-Llama-3-70B, Meta-Llama-3-70B-Instruct, Llama 3 70B

Differs in:

  • Parameters: 70B Evidence not assessed [2]Model Details table, Params

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Meta on AI Radar →