Llama-3.3-70B-Instruct is Meta's large, instruction-tuned language model designed for safe, multilingual, assistant-style conversations and text generation. With 70 billion parameters, it supports English, Spanish, French, German, Italian, Portuguese, Hindi, and Thai, offering state-of-the-art performance across a wide range of benchmarks including MMLU, HumanEval, and GPQA. The model is built on a transformer architecture with grouped-query attention, trained on over 15 trillion tokens and refined using both supervised fine-tuning and reinforcement learning with human feedback. It supports long context windows up to 128k tokens and enables advanced tool use for function calling and integration. Llama-3.3 is distributed under the Llama Community License, allowing commercial use within specific limits, and requires proper attribution and adherence to Meta's Acceptable Use Policy.
Features
- 70B parameter multilingual instruction-tuned LLM
- Supports 8 languages and 128k context length
- Built for assistant-like dialogue and advanced tool use
- Fine-tuned with SFT and RLHF for safety and alignment
- Strong benchmark performance in code, math, and reasoning
- Open weights under a custom license with attribution rules
- Safety mitigations including red teaming and refuse tuning
- Compatible with Transformers, bitsandbytes, and llama.cpp
Project Samples
