One engine. Any model. Zero Python.

One engine. NPU, GPU, or CPU. 99.89% HuggingFace architecture coverage. No Python, no runtime, no config. JARVIS — a fully local voice assistant — ships with it, out of the box.

Star on GitHub Read the story Meet JARVIS

engine online 1bp / gguf / onnx

# build from source: no installer yet
$ git clone https://github.com/1bit-MONSTER/1bit-MONSTER
$ cd 1bit-MONSTER && cmake -B build && cmake --build build
$ ./build/1bit zaya --model model.1bp   # serves on :8088
$ curl -s localhost:8088/completion -d '{"prompt":"Hello world","n_predict":32}'
$

engine census 326,992 checkpoints · 569 tokens

Every checkpoint maps. Nothing is dropped.

326,992 arch-bearing checkpoints resolve to 569 tokens, 32 families, 16 backends. Measured, not projected. And every model serves from one shared memory pool — NPU, GPU and host alias the same SharedBO pages, zero copies between lanes — with one chained API call per layer: the Windows NPU memory model, matched on Linux.

99.89% coverage / 6 hardware targets probed / 0 Python / 1 pool · 0 copies

coverage map bar width on a log scale

checkpoints326,992

architecture strings2,030

architecture tokens569

manifest families32

backend implementations12

frontier gates 17/17 green · full logits vs reference

Nemotron 3 ✓

corr 0.99986

DeepSeek V4 ✓

top-1 == HF

GLM-5.2 ✓

top-1 == HF

MiMo V2 ✓

top-1 == HF

Qwen3.5 ✓

corr 1.00000

The one gap: the CUDA backend compiles and is wired into the factory, but has never run on real NVIDIA hardware. Validation is open: testers wanted.

Build from source. That is the whole install.

Clone, cmake, run. GPL-3.0 licensed, no accounts, no telemetry.

Contribute on GitHub