incoai/splash is a local inference engine that runs language models on Apple silicon. The latest release is 1.3.2. The provided FACTS did not list any items that changed in this version. It is written in C++, Metal and Python, and uses the Apache-2.0 license.
The engine is meant for running coding agents and OpenAI- and Anthropic-compatible servers on a single Mac. It makes DFlash 2 speculative decoding the default and, alongside dedicated Metal kernels, uses the Apple Neural Engine for FFN processing. 27B models use this path with the default settings. A KV page hierarchy and a memory governor manage memory according to the model. It also supports vision, tool calling and a built-in chat page, reusing cached prefixes and batching concurrent requests.
The performance comparison is the project's own benchmark against llama.cpp, using the same Unsloth UD-Q4_K_M file on the M5 Pro and M3 Max. The document also gives the top-1 token agreement rate for a 27B model, but there is no third-party reproduction. The comparison is based on llama.cpp's default settings, so the speed advantage should not be read as a general result. The benchmark and regression scripts are public, so the measurement basis can be examined, but an independent reproduction is needed to judge performance.