Meta Unveils MobileLLM-R1 Family of Small AI Models That Run Locally on Devices for Math and Coding

Meta launches MobileLLM-R1, a breakthrough family of ultra-compact AI models with 140M to 950M parameters that excel at math and coding while running entirely on local devices, signaling a major industry shift from massive cloud-based models to specialized on-device AI that offers enhanced privacy and cost control.

• Meta releases MobileLLM-R1, a family of sub-billion parameter AI models designed for math, coding, and scientific reasoning that can run locally on devices with sizes ranging from 140M to 950M parameters
• The 950M model outperforms similarly-sized competitors on key benchmarks but is restricted by a non-commercial license, while alternatives like Google's Gemma 3 270M and Alibaba's Qwen3-0.6B offer commercially viable options
• The industry shifts toward deploying fleets of specialized small models instead of relying on single large models, offering enterprises better cost predictability, privacy control, and alignment with AI agent architectures