Lemonade Server hits Gemma 4 snag

Lemonade’s pitch is strong, but this thread is a reminder that “easy local AI” can still collapse into backend archaeology when model support lands faster than app integrations stabilize. This looks less like a single bad setup and more like Gemma 4 exposing rough edges across the macOS llama.cpp stack.

• –Lemonade positions itself as an open-source, OpenAI-compatible local AI runtime with multi-engine support, but its docs explicitly label macOS support as beta and route Apple Silicon through llama.cpp with Metal.
• –The Reddit logs show `llama-server` exiting with code `-1`, and a later update mentions `file system sandbox blocked open()`, which suggests the failure may involve macOS app permissions or binary access, not just the Gemma model itself.
• –Lemonade officially supports custom `llama-server` binaries, which is powerful but also means reliability can hinge on exactly which upstream llama.cpp build is wired into the app.
• –Related macOS reports outside Lemonade, including an April 2, 2026 LM Studio bug, showed Gemma 4 GGUF failing because bundled llama.cpp builds did not yet recognize the `gemma4` architecture, so the pain is ecosystem-wide.
• –If Lemonade keeps breaking, the lowest-friction escape hatch is usually a more mainstream local stack like Ollama or plain llama.cpp paired with Open WebUI, even if it means less integrated polish.