llama.cpp b11003 Adds Support for HRM-Text Architecture
Release b11003 of llama.cpp introduces support for the HRM-Text (DFM Mimir 1B) model architecture. The update implements custom tensor conversions, GGUF metadata handling, and looped graph building for HRM-Text's unique dual transformer stacks. This addition expands llama.cpp's flexibility, allowing users to run novel hierarchical recurrent architectures on consumer hardware. It enables local execution and testing of models designed for latent space reasoning that differ from traditional dense LLM designs. Because every token runs through 128 block passes in Mimir 1B, the decode compute cost and KV cache memory usage are approximately four times higher than a standard model of equal width. The current implementation only supports causal attention, leaving the upstream Prefix-LM mode un-implemented for now.
## BACKGROUND
llama.cpp is a popular open-source C/C++ framework enabling high-performance local inference of Large Language Models across diverse hardware architectures. HRM-Text (Hierarchical Recurrent Model) is an LLM architecture that processes text using alternating low and high transformer stacks to enhance task completion and latent reasoning efficiency.