@deemwario/mochallama

Run a local llama.cpp LLM from your terminal. The engine is JVM-powered — mochallama-core bridges llama.cpp via Project Panama (the Foreign Function & Memory API), no Python and no separate llama.cpp install required.

This npm package is a thin cross-platform launcher. The actual payload — a self-contained jlink runtime image (a trimmed JRE plus the native llama.cpp libraries) — ships in a per-platform companion package that npm installs automatically for your OS/CPU. You do not need a JDK installed to use it.

Platform support

Installs and runs on:

| platform | companion package | |------------------------|---------------------------------------| | macOS Intel (x64) | @deemwario/mochallama-darwin-x64 | | macOS Apple Silicon | @deemwario/mochallama-darwin-arm64 | | Linux (x64) | @deemwario/mochallama-linux-x64 | | Windows (x64) | @deemwario/mochallama-win32-x64 |

npm pulls in only the companion matching your host (via each companion's os/cpu fields); the others are skipped. The launcher resolves the right one at runtime.

Install / run

# One-shot via npx — drops straight into a chat session.
npx @deemwario/mochallama

# Or install globally.
npm i -g @deemwario/mochallama

Usage

# Bare invocation starts chatting immediately (defaults to `chat`).
mochallama

# List available model profiles and whether they're cached locally.
mochallama models

# Chat with a specific model (downloads it on first use into ~/.chatbot_models).
mochallama chat --model llama-3.2-1b

# Help.
mochallama --help

How it works

mochallama (this npm bin)
  └─ bin/mochallama.js                        # cross-platform launcher
       └─ @deemwario/mochallama-<platform>    # resolved via os/cpu
            └─ image/bin/mochallama           # jlink launcher (trimmed JRE + JVM args)
                 └─ mochallama-core           # Panama FFM bridge → llama.cpp natives

The jlink launcher runs with --enable-native-access=ALL-UNNAMED and --add-modules=jdk.incubator.vector baked in. A bare invocation with no arguments forwards chat so you start talking to a model right away.

Building / publishing (maintainers)

The companion packages' image/ directories are not committed — they're produced by Gradle (:cli:jlink) and copied in before publishing. Locally (macOS x64), from the repo root:

task cli:npm:pack      # builds the jlink image, stages it into the darwin-x64
                       # companion + the main package, then `npm pack` both

All four platforms' images are produced by the GitHub Actions release matrix on a v* tag — see .github/workflows/release.yml.

License

MIT