DFlash team ventures into Mac, 27B model hits 144 tokens/s.

Beating AI News Flash: Inco AI has open-sourced its local inference engine Splash, achieving up to 144 tokens/s running Qwen3.8-27B on an M5 Max MacBook Pro. LM Studio has promptly integrated it, and version 0.4.25 is already available for direct use.

Inco's previously developed DFlash has already been adopted into SGLang, vLLM, TensorRT-LLM, and llama.cpp. Meta, NVIDIA, Xiaomi, Poolside, and others have also equipped their own models with DFlash acceleration small models. DFlash works by having a small model guess multiple tokens in parallel first, then handing them to the large model for batch verification, reducing the computation of token-by-token generation.

Splash extends this optimization to the entire local inference engine. It currently only supports Qwen3.8-27B and Qwen3.6-35B-A3B, with each model individually configured with its own GPU kernel, memory scheme, and DFlash 2 small model. It covers few models, but this comes in exchange for more aggressive performance optimization.

144 tokens/s is the peak demonstrated value on the M5 Max. On the 48GB M5 Pro that Inco used for benchmarking, Qwen3.8-27B reached 74 tokens/s for a single stream with short context; with 4 concurrent streams, total throughput reached 170 tokens/s, 3.9 times that of the runner-up. This is also where Splash is better suited for Agent scenarios. When an Agent spins up multiple subtasks simultaneously, concurrent total throughput is often more important than single-conversation speed.

Splash itself is open-source and not tied to LM Studio. It currently requires an M3 or newer Mac, macOS 26.4 or above, and at least 36GB of unified memory.