70B

70B commonly refers to the size of certain large language models, where 70B stands for roughly 70 billion parameters. The label is used to describe models in the 50–100 billion parameter range, indicating a higher capacity than mid-sized models while still smaller than the largest families.

Models of this scale are typically based on the transformer architecture and are trained on large, diverse text corpora across multiple domains. The parameter count affects expressiveness, memory requirements, and potential performance, and training such models requires substantial compute resources and specialized infrastructure.

Notable examples include Meta's Llama 2 70B, among others in the same parameter range. These models can perform a variety of language tasks, such as long-form generation, summarization, translation, coding assistance, and problem solving under certain conditions. However, they can also reflect biases present in training data and may produce inaccurate or misleading results, a phenomenon known as hallucination. In practical deployment, they demand significant GPU memory and latency considerations; techniques like quantization, distillation, or offloading are often used to balance cost and responsiveness.

In AI model catalogs, 70B is one tier among several parameter-size classifications, alongside smaller models such as 7B and 13B and larger ones around hundreds of billions of parameters. The designation helps users anticipate comparative capabilities and resource requirements across models.