BERT-Large

BERT-Large is Google’s large bidirectional Transformer encoder for understanding text. BERT is trained with masked-language modeling: it learns to recover hidden words using context from both left and right, which makes its representations useful for classification, extraction, and related language-understanding tasks. It is not primarily an autoregressive chatbot that generates long answers token by token. ‘Large’ denotes the bigger standard BERT configuration, traditionally 24 Transformer layers, rather than an accelerator count. TT-Metal lists BERT-Large on a single n150 Wormhole device with batch 8 and reports sentence-oriented throughput. That one-device entry is much simpler than the multi-chip tensor-parallel configurations used by the largest generative language models in the matrix.