← All models

Llama 3.1 Nemotron 51B Instruct

NVIDIA · 2024-09-23 · 51.5B parameters

Llama 3.1 Nemotron 51B Instruct is NVIDIA's open-weight general-purpose chat model derived from Meta's Llama 3.1 70B Instruct with neural architecture search and knowledge distillation, announced on September 23, 2024; the released BF16 checkpoint contains exactly 51,501,015,040 parameters. Its 80-block DeciLM-style decoder uses NAS-selected FFN widths and varying grouped-query attention, with some blocks replacing attention by a linear layer or skipping it, so that it fits on a single H100 GPU at higher throughput than its parent. It supports a 131,072-token context window and is released under the NVIDIA Open Model License with the Llama 3.1 Community License.

Benchmark scores