AMD just shipped Lemonade SDK 10.3. The package shrank from roughly 101-107 MB down to 7-9 MB. A 10x reduction. The trick was dumping Electron in favor of Tauri, the Rust-backed framework that uses the platform's native webview instead of bundling an entire Chromium instance. Michael Larabel at Phoronix confirmed the pre-built Windows and macOS binaries saw the biggest drops. That size reduction matters. If you're running models on your own hardware, you probably care about resource efficiency. Lemonade is AMD's open-source local AI server, supporting their CPUs, GPUs, and NPUs across Windows and Linux. It competes with Ollama and LM Studio but targets the AMD ecosystem specifically, with direct ROCm integration and the ability to switch between ROCm 7.2 stable, ROCm 7.12 preview (now the default), and TheRock nightly builds. That granularity is something general-purpose tools don't offer. The 10.3 release also introduces OmniRouter, which unifies different backend engines for what AMD calls a "true omni-modal" LLM experience covering text, vision, image generation, transcription, and speech. Users can now update to specific Llama.cpp versions or enable auto-updating. Lemonade maintains OpenAI API compatibility, so integrating with existing tools should work without much fuss. One catch for Linux users: Larabel reports the AppImage build hit Wayland compatibility issues on Ubuntu 26.04 LTS during testing. AMD recommends the web UI instead. Par for the course if you've used ROCm tooling on Linux.