The DFlash team expands to Mac, with its 27B model running at 144 tokens per second.

Beating AI News (from Dongcha): Inco AI has open-sourced its local inference engine Splash, which runs Qwen3.8-27B on an M5 Max MacBook Pro at up to 144 tokens per second. LM Studio has since integrated Splash, and version 0.4.25 is available for direct use. Inco’s earlier DFlash technology has been adopted by SGLang, vLLM, TensorRT-LLM, and llama.cpp. Companies including Meta, NVIDIA, Xiaomi, and Poolside have also integrated DFlash to accelerate their small models. DFlash works by first having a small model predict multiple tokens in parallel, then submitting them to a large model for batch validation, reducing the computation required for token-by-token generation. Splash extends this optimization to its entire local inference engine. It currently only supports Qwen3.8-27B and Qwen3.6-35B-A3B, with each model having dedicated GPU kernels, memory configurations, and DFlash 2 small models. While its model coverage is limited, it delivers aggressive performance optimizations. The 144 tokens per second is the peak demo performance on the M5 Max. On the 48GB M5 Pro used by Inco for cross-device testing, Qwen3.8-27B hits 74 tokens per second for a single short-context run; with 4 concurrent runs, total throughput reaches 170 tokens per second, 3.9 times that of the second-place alternative. This is why Splash is particularly well-suited for agent use cases. When an agent launches multiple subtasks simultaneously, total concurrent throughput is often more critical than the speed of a single conversation. Splash is open-source and not tied exclusively to LM Studio. It currently requires an M3 or newer Mac, macOS 26.4 or later, and at least 36GB of unified memory.