Fermion Research Launches Neutrino-1 8B: Ternary-Format LLM
Original: Neutrino-1 8B
Why This Matters
Ternary weight compression enabling broad hardware compatibility from a single artifact could redefine LLM deployment economics.
Fermion Research released Neutrino-1 8B on July 27, 2026 — an 8.19B-parameter LLM stored in a proprietary ternary weight format, fitting in a 3.88 GB single-file artifact, 1/8th the size of fp16, scoring 72.1 on MMLU and reaching 763 tok/s on H100 with speculative decoding.
Fermion Research released Neutrino-1 8B, the flagship of its Neutrino model family, on July 27, 2026. The model is an 8.19B-parameter decoder-only transformer with 36 layers, based on Alibaba Cloud's Qwen3-8B architecture under Apache-2.0 license. Its defining feature is a proprietary 'ternary-family' weight format applied to all 252 transformer linear layers, compressing the model to a single 3.88 GB file — one-eighth the size of an equivalent fp16 model. Weights remain bit-packed at rest and are decoded inside matrix kernels at inference time, never expanding to fp16 or fp32 in memory. The compact footprint allows the model to run on an 8 GB GPU, a 16 GB MacBook, or a desktop CPU from a single artifact without conversion. A 4k-token KV cache costs 1.21 GB alongside the 3.88 GB weights. Among the 6.95B coded weights, 62.63% are zero, with near-perfect sign balance (18.68% positive, 18.69% negative). The model achieves 72.1 on MMLU and 763 tokens/second with speculative decoding on an H100. The full container is 3,875,404,812 bytes, with 67.2% allocated to transformer linears, 32.1% to token embeddings, and 0.6% to metadata.