llama.cpp b11483 Adds Support for LiquidAI d1-3B Decision Model
The b11483 release of llama.cpp adds support for LiquidAI's d1-3B decision model architecture and updates GGUF handling to read the LFM2 image resizing algorithm directly from model metadata. The internal decision type designation was also updated to lfm2-d1. Expanding support to non-generative decision models allows developers to run fast, zero-output-token decision inference directly within the llama.cpp ecosystem. This enables efficient execution of specialized multimodal classification and choice-selection tasks on edge and local devices. LiquidAI's d1-3B is a 3.12B parameter decision model that delivers structured answers in a single forward pass instead of generating autoregressive tokens. The release includes refreshed pre-built binary distributions across Windows, macOS, Linux, Android, and Snapdragon platforms.
## BACKGROUND
llama.cpp is a high-performance open-source LLM inference library written in C/C++ that enables local execution of quantized GGUF models across diverse hardware. Unlike traditional generative language models that generate text token-by-token, decision models compute direct predictions over candidate options in a single pass to achieve lower latency.