LePika v0.1.2

Fixed

• A bundled ChatML template missing from a damaged install now raises a
  FriendlyError instead of a bare FileNotFoundError. The template repair runs
  after a pull has already succeeded, so the traceback would have been the only
  thing the user saw go wrong. (#15)
• lepika down now stops an OpenWebUI that lepika up adopted rather than
  started. up returns early when the port already answers, recording no pid, and
  down had only that pid file to go on — so it reported "Nothing was running"
  while the UI kept serving and lepika status kept showing it up. It now falls
  back to the process list, the same command-line evidence a hung pid is already
  judged by. (#17)
• lepika update no longer reports "Ollama installation failed" after the engine
  upgraded fine. The official install script runs under set -e, so a machine
  whose systemd is wedged aborts it in the optional service setup — the very setup
  LePika undoes on the next line — long after the binary landed. The engine on
  PATH now decides, and the unit the aborted script left enabled is disabled as
  usual. (#18)

LePika v0.1.1

Fixed

• Repos whose weights are already quantized (NVFP4, AWQ and friends) are imported
  as-is, instead of failing after the full multi-gigabyte download because Ollama
  refuses to requantize a quantized source. The size estimate no longer promises a
  store copy a quarter the download's size. (#10)
• The systemd ollama.service that Ollama's install script enables on Linux is
  disabled after install, so it stops fighting LePika's own ollama serve for port
  11434 — and stops winning that port after a reboot with an empty model store. (#7)
• A failed Express-to-Server switch now sweeps up the containers it left behind.
  Previously they stayed invisible to lepika down while holding ports 3000 and
  11434. (#8)
• Pulled ChatML models whose published template dropped its tool handling are
  rebuilt with a tool-capable chat template. (#12)
• quant_method, read from a repo's own config.json, is escaped before it is
  printed. A value containing Rich markup was parsed as markup and silently vanished
  from the text you read to approve a download. (#13)

LePika v0.1.0

The first release of LePika: one command → local AI chat in your browser.

uv tool install lepika
lepika

Highlights

• ⚡ Express mode (default) — no Docker needed. LePika installs Ollama and Open WebUI natively (via uv), picks a model that fits your hardware, and opens chat in your browser.
• 🐳 Server mode — a docker compose stack for machines that prefer containers, with an NVIDIA overlay and a vLLM profile for full-weight models on Linux + NVIDIA.
• Model management — a curated, hardware-aware model list; refs in three shapes: Ollama tags (llama3.2:3b), GGUF pulls (hf.co/...), and full-weight org/repo imports quantized on your machine (-q nvfp4|int4). lepika model import <dir> imports weights you already have on disk.
• Remote engines — lepika connect points the UI at an engine elsewhere; lepika expose (Server mode) shares yours on the LAN behind a generated API key.
• Boring in the right places — health-checked startup, offline-friendly, generated secrets in 0600 files, every failure mapped to a one-line fix hint.

Supported: macOS (Apple Silicon), Linux, and Windows, Python ≥ 3.11.

See the README for a tour, and docs/models.md + docs/server-mode.md for the deep dives.