CLI

@deemwario/mochallama is a terminal CLI that runs a local llama.cpp model straight from the command line — a real, multi-turn local chat, in the spirit of ollama run. It is JVM-powered (Project Panama FFM bridge, the same mochallama-core engine as the Spring app) but shipped as a self-contained jlink runtime image — a trimmed JDK 22 plus the native libs — so you do not need Java installed.

One command to chat

No global install, no Java, no daemon. npx fetches the CLI, the model downloads on first run, and you're in a real conversation:

npx @deemwario/mochallama chat -m qwen2.5-1.5b

Loading model: /Users/you/.chatbot_models/qwen2.5-1.5b-instruct-q4_k_m.gguf
session a1b2c3d4
resume later with: mochallama chat --resume a1b2c3d4
Ready. Type a message, /help for commands, /exit to quit.

you> my name is Mira
bot> Nice to meet you, Mira! How can I help?
     (0.9s, ~7.0 words/s)

you> what's my name?
bot> Your name is Mira.
     (0.4s, ~3.0 words/s)

you> /exit
bye.

Real multi-turn since 0.1.6

mochallama chat keeps the full conversation history — the model sees every prior turn, so it remembers context like the name above. Earlier builds were amnesiac single-turns; 0.1.6 made chat a genuine ollama run-style session.

Install

npm i -g @deemwario/mochallama
# or run without installing:
npx @deemwario/mochallama chat

Cross-platform, no Java install

Ships per-platform packages for macOS (Intel + Apple Silicon), Linux (x86-64 + ARM64), and Windows (x86-64). npm installs only the package matching your OS/arch (via optionalDependencies + os/cpu), so you download just your platform's jlink image (~31 MB). The model itself downloads on first run.

Sessions (persistent, resumable)

Every chat is a saved session. Conversations persist as one JSON file per session at ~/.chatbot_models/sessions/<id>.json (the <id> is the 8-hex string the CLI prints on start, e.g. a1b2c3d4). The session stores the original model ref you typed and the full turn history.

List what you can resume with mochallama sessions:

mochallama sessions

ID          MODEL                 TURNS  UPDATED
--          -----                 -----  -------
a1b2c3d4    qwen2.5-1.5b          2      2026-06-05T10:14:09.220Z
9f8e7d6c    Qwen/Qwen2.5-3B-In…   5      2026-06-04T22:01:53.118Z

Resume one with: mochallama chat --resume <id>
Sessions dir: /Users/you/.chatbot_models/sessions

The TURNS column counts your user turns. Continue any prior conversation with --resume <id>. The full history is re-seeded, and the session's stored model wins — you don't repeat -m:

mochallama chat --resume a1b2c3d4

Loading model: /Users/you/.chatbot_models/qwen2.5-1.5b-instruct-q4_k_m.gguf
session a1b2c3d4
resume later with: mochallama chat --resume a1b2c3d4
Resuming session a1b2c3d4 (2 turns)
Ready. Type a message, /help for commands, /exit to quit.

you> remind me of my name
bot> You told me your name is Mira.

Want a throwaway chat that leaves nothing on disk? Pass --no-save:

mochallama chat -m qwen2.5-1.5b --no-save

In-chat slash commands

Inside the you> prompt:

Command	Effect
/help	List the available commands.
/reset	Clear the conversation history (keeps the same session id).
/exit	Quit (so does EOF / Ctrl-D).

List model profiles

mochallama models shows the built-in profiles and whether each is cached locally (in the shared ~/.chatbot_models dir, reused by the Spring app too).

mochallama models

PROFILE         FILENAME                                  SIZE       CACHED
-------         --------                                  ----       ------
qwen2.5-1.5b    qwen2.5-1.5b-instruct-q4_k_m.gguf         ~1.1 GB    no
qwen2.5-3b      qwen2.5-3b-instruct-q4_k_m.gguf           ~2.1 GB    no
qwen3-4b        Qwen3-4B-Instruct-2507-Q4_K_M.gguf        ~2.5 GB    no
phi-4-mini      Phi-4-mini-instruct-Q4_K_M.gguf           ~2.5 GB    no

Cache dir: /Users/you/.chatbot_models

These built-ins are the tool-only lineup shared with the Spring app.
Any tool-capable Hugging Face id also works, e.g.:
  mochallama chat --model Qwen/Qwen2.5-3B-Instruct-GGUF
Non-tool-capable models are refused at load.

The built-in profiles are the same tool-only set the Spring app ships (qwen2.5-1.5b is the default).

Choosing a model

You can pass any tool-capable Hugging Face id (org/repo) or a path to a local .gguf file instead of a profile name, and tune the turn:

# Built-in profile
mochallama chat --model qwen2.5-3b

# Hugging Face id — resolves the Q4_K_M GGUF and caches it
mochallama chat --model Qwen/Qwen2.5-3B-Instruct-GGUF

# Local file, with sampling tweaks
mochallama chat --model ~/models/my-model.gguf --max-tokens 512 --temperature 0.4

Options: -m/--model (profile name, HF id, or .gguf path), --max-tokens (default 256), --temperature (default 0.7), --resume <id>, --no-save. Run mochallama chat --help for the full listing.

Tool-only, enforced

mochallama only runs tool-capable models. If you pass a non-tool model (profile, HF id, or file), the CLI refuses it at load with a clear message and exits non-zero — it never silently downgrades. A gated Hugging Face model fails early too. See tool-calling support.

Shared catalogue + cache

The CLI's built-in profiles are the same tool-only set as the Spring app (qwen2.5-1.5b / qwen2.5-3b / qwen3-4b / phi-4-mini), and both share one resolver/downloader and the ~/.chatbot_models cache — a GGUF fetched by either side is reused by the other.

How it works

mochallama (npm bin)
  └─ bin/mochallama.js          # Node shim, forwards argv + exit code
       └─ image/bin/mochallama  # jlink launcher (trimmed JDK 22 + JVM args)
            └─ mochallama-core   # Panama FFM bridge → llama.cpp native libs

The launcher runs with --enable-native-access=ALL-UNNAMED and --add-modules=jdk.incubator.vector baked in. No JNI, no separate daemon, no network hop — the model runs in-process inside the CLI's own JVM image.