@jianchen1799: Local models can now handle agentic workloads. Local inference engines need to catch up. Today we’re releasing Splash: …
Summary
Inco AI releases Splash, an open-source inference engine optimized for Apple silicon, claiming up to 3× faster decode speeds for local model serving, enabling agentic workloads on devices like M5 Max MacBook Pro.
View Cached Full Text
Cached at: 09/19/26, 07:11 PM
Local models can now handle agentic workloads. Local inference engines need to catch up.
Today we’re releasing Splash: model-optimized Apple silicon serving, powered by DFlash2.
2× faster than the next-fastest engine. Qwen3.8-27B: up to 140 tokens/s on M5 Max in real agentic serving.
Inco AI (@inco_ai): Qwen3.8-27B at 144 tok/s on an M5 Max MacBook Pro ⚡
Meet Inco Splash: our open-source inference engine, built around the model and around Apple silicon.
Up to 3× the decode speed of Ollama, 2× oMLX, and almost 4× when an agent fans out into sub-agents.
Similar Articles
I've created the fastest local AI engine for Apple Silicon. Optimised for agentic use.
The author announces the release of 'lightning-mlx', a local AI engine optimized for Apple Silicon that achieves high token speeds for coding agents and tool-calling workflows.
@rohanpaul_ai: Massive move by Apple for local inference. They just launched M6 and M5 Ultra to move more AI inference onto Mac deskto…
Apple has launched the M6 and M5 Ultra chips to enable advanced local AI inference on Mac desktops, with high unified memory and compute specifications supporting large language models on-device.
@svpino: DeepSeek-V4-Flash running at 5.71 token/s on a Mac M5 Pro. Every day, we get better models running on consumer hardware…
The tweet highlights DeepSeek-V4-Flash running at 5.71 tokens per second on a Mac M5 Pro, emphasizing advancements in local AI inference on consumer hardware, with a mention of Tencent's open-source Palm-Infra for Apple Silicon optimization.
Localmaxxing (3 minute read)
The article analyzes the viability of running AI inference locally on a MacBook Pro, comparing a local Qwen 35B model against the cloud-based Claude Opus 4.5. It concludes that local models are 2x faster for routine tasks, making them a practical choice for half of daily workloads despite a slight capability gap.
Splash on a 40-core M5 Max: +20% decode by tuning the kernels for your own chip
The article describes how to use the 'make tune-kernels' tool in Splash to optimize AI model decoding on 40-core M5 Max chips, achieving up to 20% speed improvement by tuning kernel layouts for specific hardware.