Fermion Research Unveils Neutrino-1 Model Family

Fermion Research has unveiled a family of 3 Neutrino-1 models that directly apply a Ternary-style format starting from the training stage. Rather than the conventional post-training quantization approach, these models are designed via Native-format training, where all forward passes occur within the low-resolution format.

Key Model Lineup:

• Neutrino-1 8B: An 8B-class flagship model (MMLU 72.1)
• Neutrino-1 0.6B: A draft model and small model
• Neutrino-1 0.6B-Chat: A chat-specialized model

Core Technology and Features:

• Storage and Bandwidth Savings: The 8B model is highly efficient, with a 3.88GB disk footprint and a 2.56GB download size, addressing the bandwidth bottleneck that arises during single-stream decoding.
• Unified Container Architecture: All three models share the same tokenizer, container layout, and inference engine, allowing the 0.6B model to be used immediately as a Draft model for the 8B model.
• GQA (Grouped-Query Attention) Applied: An efficient attention mechanism has been adopted.
• Open Source: All models are licensed under Apache License 2.0 and are readily available on Hugging Face.