8.19B; 596M (0.6B and 0.6B-Chat)
40,960 tokens
Apache 2.0 weights; bundled optimized binaries have separate proprietary terms
Grouped release of three compact open-weight language models trained for Fermion Research's proprietary ternary-family representation: the general-purpose Neutrino-1 8B chat model, a Neutrino-1 0.6B base and speculative-drafting model, and the conversationally post-trained Neutrino-1 0.6B-Chat. The 8B model is derived from Qwen3-8B through ternary quantization-aware training and staged post-training for knowledge, instruction following, mathematics, tool use, and clean termination. It contains exactly 8,190,735,360 parameters across 36 decoder layers, with a 4,096-dimensional hidden state, 12,288-dimensional gated SwiGLU feed-forward blocks, 32 query heads, eight key-value heads, a 151,936-token vocabulary, and a 40,960-token context window. Its 252 attention and feed-forward projection matrices contain approximately 6.95B ternary weights, while its separate input and output embedding tables remain int8 and its normalization tensors remain full precision. The serving container occupies approximately 3.88 GB and its losslessly compressed transport download is approximately 2.56 GB, compared with roughly 16 GB for an equivalently sized FP16 checkpoint. The two 0.6B variants contain exactly 596,049,920 parameters and retain the same tokenizer and context length in a narrower 28-layer architecture with a 1,024-dimensional hidden state, 3,072-dimensional feed-forward blocks, 16 query heads, eight key-value heads, and tied int8 embeddings. The base 0.6B can generate independently or draft tokens for the 8B without vocabulary remapping; its repository was updated on July 31 with a draft-specialist checkpoint distilled from 8B rollouts, while the original launch checkpoint remains available through revision pinning. Fermion reports that greedy speculative decoding preserves the 8B's output exactly, with a launch certification covering 27,648 generated tokens without divergence. Published launch evaluations for the 8B include 72.1 on five-shot MMLU, 67.8 on generative MMLU-Redux, 77.2 prompt-strict on IFEval, 68.9 macro across thirteen BFCL v3 subsets, and 53.4 on zero-shot GSM8K using flexible extraction. Fermion reports 396 tokens per second for plain decoding and up to 763 tokens per second with the 0.6B draft on an H100, plus local inference through a shared Python loader, MLX pack, public llama.cpp fork, GGUF-compatible custom tensor format, and bundled x86 and Apple-silicon runtimes. The weights are Apache-2.0, but Fermion's optimized prebuilt runtime binaries remain closed-source and carry separate redistribution and reverse-engineering restrictions.