Llama 3.1 8B Instruct

Meta Platforms · United States
Per current documentation, Llama 3.1 8B Instruct is released under the Llama 3.1 Community Licence — a custom source-available licence rather than OSI open source. Commercial deployment is permitted below 700M MAU subject to the Acceptable Use Policy and attribution rules, but training-data opacity and US origin create EU AI Act transparency and data-transfer gaps that deployers should document.
Licence facts
Parameters
~8B
Architecture
Decoder-only transformer with Grouped-Query Attention, SFT + RLHF instruction tuning
Context length
128K tokens
Languages
English, German, French, Italian, Portuguese, Hindi, Spanish, Thai
Knowledge cutoff
December 2023
Released
2024-07-23
Known risks
  • Licence restrictions travel with every downstream fine-tune: the AUP field-of-use carve-outs (weapons, CSAM, critical infrastructure, military/ITAR, mass surveillance, unlawful discrimination, malware, disinformation) and the 'Built with Llama' attribution rule must be enforced contractually with integrators.
  • Training-data disclosure is a headline figure only (~15T tokens from public sources, >25M fine-tuning examples including synthetic data) with no dataset composition, sources, or filtering criteria — thin for EU AI Act Art. 53(1)(d) 'sufficiently detailed summary' scrutiny.
  • US controller exposure: self-hosting avoids CLOUD Act reach for the weights themselves, but any fine-tuning on EU personal data shifts full controller obligations (lawful basis, DPIA, records) onto the deployer, and Meta's own EU training-data practices have drawn regulatory scrutiny.
Reviewed by Ali Madjaji · Last reviewed 2026-04-17· Reviewed 1 day agoSuggest a correction