User · Quick start

Getting started

The engine is one binary. Build it, point it at a model, and it runs on NPU, GPU, or CPU with no config, no Python, and no glue. Here is the shortest path from nothing to a model.

What you get

build/1bit is a single MIT-licensed C++23 binary. It is the engine, the serving server, the voice assistant, and the CLI all at once. There is no virtualenv, no pip, no runtime interpreter. The repo's one through-line is the engine (NPU + GPU + CPU) and the app that proves it: JARVIS.

One engine, any model, zero Python. The routes from checkpoint to kernel path are detected at load, not configured.

Install

Build from source with the bundled installer. The script detects the platform, pulls header-only deps, and warns on a known-bad kernel.

$ git clone https://github.com/1bit-MONSTER/1bit-MONSTER && cd 1bit-MONSTER
$ bash install.sh            # build + nothing else
$ bash install.sh --with-jarvis  # build + JARVIS launcher + config

That is the whole install. Source to a running binary in three commands, and the app adds the voice pipeline when you ask for it.

Run a model

Point it at any supported checkpoint. Backend selection and kernel routing are automatic.

$ ./build/1bit zaya                    # serve on :8088
$ ./build/1bit zaya --port 8080        # custom port
$ ./build/1bit zaya --model /path/to/model.h1b

Or skip the server and sit in the voice assistant, which is the reference app that exercises the whole pipeline in one process.

$ ./build/1bit jarvis --model "Qwen3-0.6B" --text        # text chat
$ ./build/1bit jarvis --model "Qwen3-0.6B" \
    --whisper models/whisper-tiny.gguf \
    --piper-model ~/piper/en_US-lessac-medium.onnx    # voice

First health check

Once the server is up, hit the health endpoint before you trust anything.

$ curl http://localhost:8088/
# {"status":"ok","model":"Zaya1-8B",...}
$ curl -X POST http://localhost:8088/completion -H 'Content-Type: application/json' \
  -d '{"prompt":"The future of AI is","n_predict":32}'
# {"tokens":[...],"text":"...","gen_ms":2850.12,"tok_s":11.2}

What's next

• Build from source on your hardware, including the NPU and optional GPU decode, in the building guide.
• Serve it behind the OpenAI-compatible API in running & serving.
• Pick a model from the supported families in model families.
• Talk to it with the reference voice app in JARVIS.