Neutrino-1 8B (www.fermionresearch.com)

Fermion Research has announced the release of the Neutrino-1 8B model, an advanced 8.19 billion-parameter decoder-only transformer designed for efficient multi-platform usage. The model is notable for its proprietary ternary-family weight format, allowing it to store data eight times smaller than traditional fp16 formats, effectively optimizing memory use and decoding speeds. The entire model is encapsulated within a single 3.88 GB file, providing seamless integration across datacenter GPUs, Apple silicon, and desktop CPUs without the need for conversion. This efficient architecture enables quicker decoding rates—up to 33.7 tokens per second on Apple M5 devices—making it accessible for a variety of computational environments. The architectural innovations in Neutrino-1 include grouped-query attention that minimizes the Key-Value (KV) cache size while maintaining robust performance. By featuring a dynamic controller that sizes token drafts intelligently, the model achieves an impressive accuracy of 96.5% on factual prompts, further enhancing its utility. As an open-source tool under the Apache License 2.0, Neutrino-1 encourages commercial use and modification, positioning itself as an accessible, versatile solution in the rapidly evolving AI/ML landscape. This model promises to empower developers and researchers by providing high-efficiency, state-of-the-art capabilities across multiple platforms.