Meta Open-Sources New AI Model for On-Device Reasoning

On 22 September 2025, Meta announced the open-sourcing of its MobileLLM-R1 series of on-device models, designed to run directly on mobile devices and optimised for reasoning in mathematics, programming, and scientific domains. MobileLLM-R1 is the newest addition to the MobileLLM family, with base models and supervised fine-tuned variants at 140M, 360M, and 950M parameters; the base models handle 4K context, while the fine-tuned versions extend this to 32K tokens.

Meta emphasises that MobileLLM-R1 is not a general-purpose chatbot but is engineered for targeted reasoning tasks, including code writing in languages such as Python and C++, and mathematics and scientific challenges. According to Meta, the 950M model trained on less than 5TB of high-quality data (with only 2TB of pretraining tokens) outperformed Qwen 3 0.6B, trained on 36TB of data, across benchmarks like MATH, GSM8K, MMLU and LiveCodeBench.

The models are released on Hugging Face under the Apache 2.0 licence and can be run with the vLLM inference engine by registering the architecture as Llama4ForCausalLM in the ModelRegistry, enabling developers to build edge AI applications at lower costs without full cloud dependence.