Splash is an open-source inference engine by Inco AI designed to run large language models on Apple silicon. The engine utilizes Metal compute kernels and speculative decoding via DFlash 2 to accelerate token generation. It supports OpenAI and Anthropic API compatibility, allowing integration with various clients.
The engine requires macOS 26.4 or later and hardware with an M3 chip or newer. Memory requirements vary based on the model and quantization; 4-bit examples require at least 36 GB of unified memory, with 48 GB recommended. Compact GGUF variants can run on systems with 24 GB of memory.
Splash supports specific GGUF builds and 4-bit MLX models, including Qwen3.8-27B and Qwen3.6-35B-A3B. Performance depends on the specific model, quantization, and context reuse. The engine is licensed under Apache 2.0, with model weights subject to their own respective licenses.
Benchmarks provided by the developers for Qwen3.8-27B on an M5 Pro chip show performance gains compared to llama.cpp. However, the engine requires specific support for each model family, meaning new architectures cannot be added arbitrarily.
