Developers introduced Splash to accelerate Qwen on Apple Silicon

The team behind Splash has released a local inference engine optimized for Apple Silicon chips. The tool is designed to run Qwen family models and provides a 2-fold performance boost compared to standard execution methods.

High-speed operation is achieved through the use of speculative decoding techniques. The engine allows deploying a local server that accepts requests via API compatible with OpenAI and Anthropic standards. By default, the service runs on port 8000, simplifying integration with existing client applications and libraries.

In internal performance tests, the system demonstrated a score of 10 points when performing tasks under 5 hours of continuous load. The software is specifically designed for the Apple Silicon architecture, maximizing the capabilities of neural cores and unified memory of Apple processors.

The solution enables developers to run modern LLMs locally without relying on cloud APIs, while maintaining a familiar interaction interface. Using Qwen with the optimized engine reduces latency when generating text, which is critical for interactive AI tools and local agents.