Inco AI Releases Splash to Double Qwen Speeds on Mac

Inco AI has introduced Splash, an open-source inference engine under the Apache-2.0 license designed specifically to optimize local large language model execution on Apple silicon. In benchmark tests conducted on an M5 Pro system equipped with 48 GB of unified memory, Splash ran the Qwen3.8-27B model at 74 tokens per second. This performance is twice as fast as the next-best engine, oMLX, which clocked in at 38 tokens per second under the same conditions.

To achieve these speeds, Splash utilizes fused Metal kernels, a DFlash 2 draft model, and an 8-bit KV cache tailored for each model. It packages supported models with 4-bit weights, a matching draft model, and a fixed memory plan. For developers running complex workflows, the engine's prefix caching system dramatically reduces latency. On 32K prompts, cache reuse allows Splash to return the first token in just 282 milliseconds, representing a 7.3-fold speedup compared to oMLX.

The engine is specifically optimized for coding agents that frequently process expanding conversation histories and share repository contexts. When handling four concurrent subagents, Splash achieved a 3.9-fold aggregate decode throughput advantage over competing engines like oMLX, Lily, uzu, and Ollama. This evaluation utilized coding prompts from NVIDIA's SPEED-Bench, with output capped at 1,024 tokens.

For practitioners looking to deploy the tool, Splash requires an Apple M3 chip or newer, macOS 26.4, and a minimum of 36 GB of unified memory, though 48 GB or more is recommended. Installation is streamlined via Homebrew using the command "brew install incoai/tap/splash". Serving the 17.4 GB Qwen3.8-27B-Splash model package automatically initializes a local server on port 8000. The release also includes built-in client helpers such as splash opencode, splash claude, splash codex, and splash hermes to simplify integration.