Open weights, closed format

Fermion Research releases Neutrino-1 8B, an ultra-lightweight model for laptops whose proprietary format blocks its use on Ollama and LM Studio.

Fermion Research is releasing three models under the Apache 2.0 license: Neutrino-1 8B and two variations with 0.6 billion parameters. The selling point lies in the size: the 8.19 billion parameters of the largest model fit into a 2.56-gigabyte download and 3.88 gigabytes on disk, whereas an equivalent 16-bit model weighs sixteen. The exact same file, down to the byte, loads onto an H100 card just as it does on a 16-gigabyte MacBook, served by a single engine.

The format places each of the 252 projection matrices into a ternary family, where each weight is reduced to a choice between minus, zero, and plus, accompanied by a scaling factor per row. In this format, 62.6% of the weights are exactly zero. Vocabulary tables escape this treatment and remain in int8, accounting for nearly a third of the file. The company specifies that the extensions it brings to this family are not published.

The small model serves as a draft for the large one. It proposes up to seven tokens that the 8B validates in a single speculative pass, keeping only the prefix on which both agree. Fermion claims 763 tokens per second on an H100 for the most predictable outputs, compared to 396 without this mechanism, and publishes an unusual verification in this domain: 27,648 consecutive tokens produced identically with and without the draft.

Two caveats emerge from a cross-reading of the documents. The announcement asserts that the model was trained natively in its distribution format, without a full-precision version rounded after the fact. Meanwhile, the technical datasheet published on Hugging Face describes it as a derivative of Qwen3-8B obtained through quantization-aware training, followed by post-training. The two formulations do not align. Furthermore, this same datasheet reports lower scores for tool use and instruction following than those in the announcement: 65.31 and 73.17 compared to 68.9 and 77.2. Finally, the weights are open, but the format is not: the GGUF package only loads in the modified version of llama.cpp published by the creator, and neither in the original llama.cpp, nor in Ollama, nor in LM Studio.