Neutrino-1 8B

• Every one of its 252 transformer linears is stored in a proprietary ternary-family weight format eight times smaller than fp16; the weights stay bit-packed at rest and are decoded inside the matrix kernels, so nothin