Llama 3.1: 405B open weights, long context, tool calling
Meta released Llama 3.1 in 8B, 70B and 405B sizes with multilingual support and long context, plus Llama Guard 3 and Prompt Guard; Hugging Face provided FP8, AWQ and GPTQ quantised versions.
What happened
Meta released Llama 3.1: eight open-weight models on the Hub (3 base, 5 fine-tuned), with Hugging Face integration across Transformers, TGI, HuggingChat (405B instruct) and several clouds.
The technical details
From the Hugging Face post:
- three sizes: 8B (consumer GPU), 70B (large-scale apps) and 405B (for synthetic data, LLM-as-a-judge or distillation)
- multilingual, with long-context support
- built-in and custom tool calling in the prompt format
- safety models: Llama Guard 3 (classifies inputs and outputs) and Prompt Guard (detects prompt injections and jailbreaks)
- quantised FP8, AWQ and GPTQ versions for easier inference, plus memory-requirement tables for inference and training
Why it matters
A 405B open-weight model changed what “open” could mean at the top end, mainly as a teacher for smaller models. Prompt Guard is an early example of the layered defences I described in my agent-safety post.
Sources
This is a summary written from the linked public sources. Figures and numbers are as reported there.