Llama-3.3-70B-Instruct

Meta released Llama 3.3 on December 6, 2024, as documented on the official Hugging Face model card for meta-llama/Llama-3.3-70B-Instruct. The card confirms Meta Platforms (Ireland Limited for EEA/Switzerland; Inc. elsewhere) as the licensor and creator of the underlying foundational model and weights. Access to the gated model card requires agreeing to share contact information under the Meta Privacy Policy before the Llama 3.3 Community License terms apply. Use of the Llama 3.3 Materials is governed by the Llama 3.3 Community License Agreement, which grants a non-exclusive, worldwide, non-transferable, royalty-free limited license to use, reproduce, distribute, and create derivative works, subject to Meta's Acceptable Use Policy and redistribution obligations. The card explicitly names Meta as publisher and directs licensees to https://www.llama.com/docs/overview for documentation and https://www.llama.com/llama-downloads for distribution of the weights and code, reinforcing Meta's sole creator attribution for this model family.

The NVIDIA NGC catalog entry confirms Llama 3.3 70B Instruct as Meta's pretrained and instruction-tuned 70B text-in/text-out multilingual model, with Model Developer listed as Meta and a release date of December 6, 2024. The page notes that inference on NVIDIA GPUs is accelerated by TensorRT-LLM and that the model is approved for commercial use under the Llama 3.3 Community License Agreement with Built with Llama terms. Status is described as a static model trained on an offline dataset. Architecturally, NVIDIA's catalog summarizes Llama 3.3 70B as an auto-regressive transformer that uses supervised fine-tuning and reinforcement learning with human feedback to align with human preferences for helpfulness and safety. All model versions employ Grouped-Query Attention for improved inference scalability. The entry disclaims NVIDIA authorship, attributing development to Meta and pointing to the upstream model card for full details on the underlying weights and training.

Llama 3.3 70B Instruct is Meta's instruction-tuned, text-only multilingual large language model released on December 6, 2024. According to the InferenceX technical overview citing the Hugging Face model card, it is an auto-regressive optimized transformer with 70B parameters (the Hugging Face checkpoint lists 71B), a 128k context window, and pretraining on more than 15 trillion tokens with a December 2023 knowledge cutoff. The model officially supports English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai, and uses Grouped-Query Attention for inference scalability. Meta positions Llama 3.3 70B as delivering performance similar to the much larger Llama 3.1 405B at a fraction of the serving cost. InferenceX, drawing on Meta's instruction-tuned comparison table, reports Llama 3.3 70B matches Llama 3.1 70B on MMLU at 86.0, scores IFEval 92.1 versus 88.6 for the 405B, and reaches MATH 77.0 versus 73.8. The model is distributed under the Llama 3.3 Community License Agreement, which requires Built with Llama attribution and a separate license for licensees exceeding 700 million monthly active users.