Chat in your terminal

Lemonade ships with a built-in chat REPL so you can talk to a local model without leaving the terminal, no browser, no editor integration, no extra dependencies.

lemonade chat

That's it. If no model is loaded, you'll be prompted to pick one. If a model is already loaded server-side, it's used automatically.

Launching

Two equivalent entry points:

# Open the REPL, optionally with a model
lemonade chat
lemonade chat Qwen3-1.7B-Hybrid

# Or piggyback on the existing run command
lemonade run Qwen3-1.7B-Hybrid --chat-cli

lemonade chat is the recommended form. lemonade run --chat-cli exists so muscle memory from other tools still works.

Options

Using the REPL

Once you're in, the prompt is just >. Anything you type that doesn't start with / is sent as a user message. Streaming output flows back inline.

  ─── Qwen3-1.7B-Hybrid ───────────────────────────────────  ? /help for shortcuts
  Clear line: Esc / Ctrl-U · Stop response or exit: Ctrl-C

  >

Type /help at any time for the full list of slash commands (switching models, toggling multi-line input, viewing stats, listing or unloading models, etc.).

Reasoning models

If the active model emits a thinking trace (Qwen 3, Phi-4-mini-reasoning, DeepSeek-R1, …), the REPL renders it above the answer:

The reasoning trace is display-only, it isn't kept in the chat history sent back on the next turn.

Use /think to control whether the model reasons at all:

Key bindings