Inco AI has open-sourced Splash, a local inference engine for Mac, and said Qwen3.8-27B reached a peak of 144 token/s on an M5 Max MacBook Pro. LM Studio moved quickly to add support, and the engine is already available directly in version 0.4.25.
From DFlash acceleration to a full inference engine
Inco’s earlier DFlash work has already been integrated into SGLang, vLLM, TensorRT-LLM, and llama.cpp. Meta, NVIDIA, Xiaomi, and Poolside have also paired DFlash with their own models to speed up smaller models.
The method works by having a smaller model predict multiple tokens in parallel, then handing those predictions to a larger model for batch verification. That reduces the compute burden of generating output one token at a time. Splash takes that same idea and extends it across the full local inference engine.
Model support is narrow, but heavily tuned
At this stage, Splash supports only Qwen3.8-27B and Qwen3.6-35B-A3B. Inco said each model has its own GPU kernel, memory scheme, and DFlash 2 small model. The trade-off is limited model coverage in exchange for more aggressive performance tuning.
Benchmark figures on M5 Max and M5 Pro
The 144 token/s figure is the highest demo result Inco showed on M5 Max hardware. On the 48GB M5 Pro system the company used for side-by-side testing, Qwen3.8-27B reached 74 token/s in a single short-context stream. Under four-way concurrency, total throughput rose to 170 token/s, which Inco said was 3.9 times the second-place result.
That is also why Splash is presented as a better fit for Agent workloads. When one Agent spins up several subtasks at the same time, aggregate concurrent throughput can matter more than the response speed of a single conversation.
Open-source release and system requirements
Splash is open-source and does not require LM Studio. Current requirements are a Mac with an M3 chip or newer, macOS 26.4 or later, and at least 36GB of unified memory.