ZeroDegress/laya-gguf

Laya — GGUF

GGUF version of convaiinnovations/laya, packaged for the laya-rust inference engine.

Laya is a non-autoregressive System One decision model: 421M parameters, a ModernBERT-large encoder plus a decision head. Give it a state (text, email, ticket, JSON) and typed questions; it returns typed answers with probabilities in a single forward pass. It never generates text.

Files

file	weight type	bytes	sha256
laya-f16.gguf	F16	848,158,048	62db2affc0fe4b9f2538ba61299e5fd8bf876f50f765089a343cd4113eef21a4
laya-f32.gguf	F32	1,690,745,696	63429326ecfe75faff55b2a637b67a3012603057532cc7ebffdbc308798abf16

Both hold the same values — the published checkpoint is F16, and the F32 file is a lossless widening of it — and they return identical answers. Each file is self-contained (weights, architecture, tokenizer, fitted calibration temperatures), so no sidecar files are needed.

Run it with laya-rust

git clone https://github.com/zerodegress/laya-rust && cd laya-rust
cargo build --release --no-default-features --features mlx     # Apple Silicon; cuda is the default feature

./target/release/laya test -m /path/to/laya-f16.gguf - < request.json

This is not a llama.cpp GGUF: the architecture is laya, which llama.cpp does not know.

Attribution

Weights: Convai Innovations and Laya contributors, Apache-2.0 — see convaiinnovations/laya. Converted with the engine's own laya convert, and verified tensor by tensor against the published checkpoint.