Why even powerful M1 Max and Ultra machines are left out
The immediate bad news is that a large memory pool does not save an M1 Max or M1 Ultra from Splash's restriction. The current released build requires Apple M3 or newer, macOS 26.4, and at least 36 GB of unified memory. Inco's requirements description already recommends 48 GB.
The key barrier is hardware-related, not simply a matter of capacity. Community reports indicate that the engine stops on older Macs with a GPU-family error. The required GPU family 9 begins with M3, so a machine with plenty of memory still fails the compatibility check.
This is where Splash becomes interesting: it is not designed as a universal runtime for arbitrary models. It is a specialized engine with Metal kernels and preconfigured model layouts optimized for specific targets. That approach can extract more speed from supported hardware, but it makes compatibility part of the architecture rather than a setting that can easily be disabled.
For models in the 27B class, unified-memory capacity, memory bandwidth, and available GPU features all matter at once. The M3 restriction therefore looks less like a cosmetic filter and more like a consequence of the chosen optimization strategy. Current official compatibility does not let you simply move a Splash package to M1 or M2.
What remains for older Apple Silicon owners
For M1 and M2 users, the practical route is currently straightforward: use a more universal engine and select a model based on the memory available. LM Studio with llama.cpp and Metal supports M1, M2, M3, and M4 generations, while running llama.cpp directly provides the same basic path without the interface.
The trade-off is clear. Instead of Splash's optimizations, you may need a smaller GGUF model or more aggressive quantization. Speed and quality will depend not on the attractive chip name, but on the specific model, its format, and how many layers actually fit in unified memory.
I would first check GPU-family support in the chosen runtime rather than the advertised RAM capacity. The Splash story illustrates a new boundary in local AI: an older Mac may still have enough compute power, yet be too old for a particular set of Metal kernels.
This is not the end of M1 Max and Ultra for local models. It is the end of the assumption that a large unified-memory pool automatically guarantees compatibility with every new specialized engine.