GGUF Libra is an agent architecture that addresses context limitations in traditional dialogue systems through a tree-based conversation model with Git-like branching, spectral graph analysis for reasoning depth quantification, and local inference via llama.cpp. It replaces linear context with version-controlled branches, uses graph Laplacian eigenvalues to prioritize retrieval, and provides explicit KV cache management on GPU hardware.
Lichen is an open-weight local model system that replaces Jev for decision-making tasks with image support, running on consumer hardware via vLLM. It achieves comparable performance to Jev (207/231 on JevBench) with 56ms median latency and accurately interprets images, including identifying road signs with 92% accuracy.