benmaster82/pithos is a dependency-free C11 inference engine that runs Liquid AI's LFM2 and LFM2.5 models. The repository provides terminal chat and an OpenAI-compatible server, together with a browser-based Pandora agent and a terminal agent, and presents a setup that packs these features into a single executable. There is no separate recent release information. The license is MIT.
An inference engine is a program that takes input and generates output from a model. pithos is written in C11 without external libraries and can optionally use OpenCL acceleration. The repository publishes the code for model processing and quantization, CPU SIMD, the server and the tools, along with test files. So it does not stop at offering a single executable; the code and tests that make up LFM2 inference can also be examined. The README presents a 1.9x generation speed and a 3 to 10 second server startup, compared with the CUDA implementation of llama.cpp on a GT1650 laptop.
The performance figures, however, are experimental results presented by the repository author, and cannot be read as comparisons verified inside the repository itself. The star count is 3, and the published comparison alone makes it hard to judge that the same performance would appear in other environments. The composition of pithos is worth looking at, but the evidence is limited for deciding on adoption from performance numbers alone.