Mirai Raises $10M for On-Device AI on Apple Silicon
AI startup Mirai has announced a $10 million funding round to accelerate its on-device AI inference technology for Apple Silicon. The company claims its framework delivers up to 37% faster performance than Apple's own MLX framework. This highlights a competitive and rapidly developing ecosystem for AI optimization on Apple's hardware.
- The founding team consists of Dima Shvets and Alexey Moiseenkov, who previously created the viral AI applications Reface and Prisma, which had over 300 million and 100 million users respectively. Their experience with the performance challenges of on-device AI in consumer apps directly informs Mirai's strategy. - The $10 million seed round was led by Uncork Capital, whose managing partner, Andy McLoughlin, highlighted the unsustainable economics of cloud-based inference as a key driver for investment in edge solutions. Other backers include the co-founder of ElevenLabs and an early pioneer from Snowflake. - Mirai's proprietary inference engine, built in Rust, is engineered to be a hardware-aware execution layer that optimizes the interaction between AI models and the underlying silicon. The company is advised by a former Apple Distinguished Engineer who worked on the MLX framework. - Beyond the 37% faster generation speed, the company also reports up to a 59% improvement in prefill performance on specific model-and-device combinations when compared to both Apple's MLX and llama.cpp. - Mirai's product aims to abstract away the complexities of on-device AI, such as memory management and kernel optimization, allowing developers to integrate local inference with just a few lines of code. - The company's roadmap includes expanding support beyond Apple Silicon to other platforms and extending its capabilities to handle voice, vision, and multimodal AI workloads. - To manage workloads that are too large for local processing, Mirai is developing an orchestration layer to intelligently and automatically route tasks between the on-device engine and cloud-based models.
