The terminal gets a seat

Quenderin now runs where programmers live. quenderin chat is a streaming REPL in your terminal — same model catalog, same SHA-256-verified downloads, same zero-network promise as the apps. And because it's a proper CLI, it composes:

git diff | quenderin chat -p "review this change"
cat error.log | quenderin chat -p "what's the likely cause?"
quenderin chat -m qwen25-coder-7b        # the coder model, interactively

That first line matters more than it looks: it's code review that never leaves your machine. Proprietary diff, client code, unreleased work — piped into a model that has no network to leak it to.

What it is

Measured, not promised

On the 10-core Apple-silicon laptop we build on — deliberately tested while the machine was starved, with under 3 GB of RAM free — the 0.4 GB ultra-light model streamed at 78–101 tokens/second with ~90–150 ms to first token, after the engine auto-tuned its context window down to fit. Bigger machines get bigger models; the CLI makes the same honest RAM-fit calls as the apps.

The two bugs we hit (kept in public, as usual)

First: chat initially failed before the first token on that starved machine. The shared chat path prepended a ~465-token tool preamble to every session, and the auto-shrunk 512-token context couldn't fit it. The CLI now runs a plain-chat session — and the failure mode is in the public ledger.

Second: after a perfect answer, the process exited with code 134 — a known llama.cpp Metal teardown assertion. Harmless-looking, fatal in practice: a pipeline checking exit codes would treat every successful answer as a failure. The fix is a full engine dispose before exit. If your scripts wrap AI tools, check your exit codes.

Get it

It ships with the source today — npm install && npm run build:tsc && npm link from the repo puts quenderin on your PATH. A standalone npm package is planned. Where this fits in the bigger picture: the roadmap.

← Back to Quenderin · Blog · Changelog · Download