Neutrino-1 8B

Neutrino-1 8B is an 8.19 billion parameter language model compressed to 3.88 GB using a proprietary ternary-family weight format, eight times smaller than standard fp16 encoding. The compressed format enables efficient deployment on consumer hardware like 8 GB GPUs or 16 GB laptops while maintaining fast single-token inference speeds that exceed larger uncompressed models on the same systems. The model uses grouped-query attention and achieves approximately 63% sparsity across its weights, with the compression applied to 252 transformer linear layers while embeddings and normalization weights remain in int8 or float32 formats.