v2026.6.4
Alenface is not just another LLM wrapper; it’s a high-performance native macOS client built for the next generation of
efficient AI. While other local LLM tools are often built on bloated frameworks and consume gigabytes of space, Alenface
is engineered for speed and precision.
How did we do it? Pure Java (25) powered by Llama.cpp and integrated with a native browser engine. No Electron. No
Chromium. No JCEF. No compromises.
Key Highlights:
Ternary / 1-bit models (Bonsai): Full support for the PrismML Bonsai-8B family via the prism llama backend
(libllama 0.0.8846, ggml 0.9.11). Run high-quality 8B-parameter intelligence from a model file of only ~1.24 GB on disk.
Hugging Face Integration: Search, download, and manage GGUF model variants directly within the app. The model list shows
parameters and size at a glance; per-variant quant size is resolved on demand when you open a card, so multi-quant
repositories report the exact size of the selected file.
Native Performance: Built on Apple Silicon with a system WebView (WebKit) for the UI and a JNI bridge to llama.cpp for
inference — a single process, no Chromium, no network IPC. Resource monitoring goes through the native
libalenmonitor.dylib helper.
Extreme Efficiency: A compact application distribution (DMG, not JAR), an order of magnitude lighter on memory than
Electron-based clients.
Engineering notes (this release):
- Toolchain moved to JDK 25; runPom.xml fixed for Java 25 (preview flags removed, JNA 5.14.0 added, JavaFX 23.0.2).
- Hugging Face API repair: query URL-encoding (multi-word search no longer fails); model size from gguf.totalFileSize,
parameter count from gguf metadata; per-variant size via tree/main?recursive=true. - Test suite: 44 network-free, mock-free JUnit 5 tests across model parsing, HF service, inference params, MD generation,
number formatting, and the download manager.
Bonsai: We are closely monitoring the evolution of these remarkable models. Our mission is to provide continuous support
for this architecture, ensuring that anyone can run the most cutting-edge and efficient AI solutions directly on a
standard laptop.