As generative AI moves towards lightweight and localized development, Meta announced the open source release of the latestMobileLLM-R1 Series Models, which can be deployed and run directly on mobile devices and other devices, and focuses on reasoning and processing of mathematical, programming, and scientific problems. This move not only responds to market demand for "on-device AI," but also demonstrates the continued investment of major technology companies in improving reasoning capabilities.
MobileLLM-R1 is a new member of the Meta MobileLLM family, distinguished by its streamlined and specialized features. The series includes two types of models: a base model and a final model after supervised fine-tuning (SFT), with parameters of 1.4 million, 3.6 million, and 9.5 million, respectively. The base model supports a context length of 4K tokens, while the final model can be expanded to 32K tokens, significantly improving its ability to handle complex problems.
Meta emphasized that MobileLLM-R1 is not a general-purpose language model for chatbots, but is designed for specific reasoning scenarios, such as math problem solving, programming (including languages like Python and C++), and scientific research-related tasks.
雖然其最大版本MobileLLM-R1 950M僅以不到5TB高品質資料完成訓練 (其中預先訓練資料僅2TB組token)，但表現仍相當驚豔。Meta表示，在MATH、GSM8K、MMLU、LiveCodeBench等多項評測中，MobileLLM-R1的成績超越使用36TB組訓練資料的Qwen 3 0.6B模型。
In a more detailed comparison, the MobileLLM-R1 950M's accuracy in the MATH test is five times that of Olmo's 1.24B and twice that of SmolLM's 1.7B. It also leads in code generation and problem-solving capabilities.
而更小的MobileLLM-R1 140M (base)也優於SmolLM2-135M，至於360M版本更以大幅差距超過Gemma-3-270M-pt與SmolLM2-360M (base)，凸顯Meta 在模型架構與訓練策略上的最佳化成果。
It is worth noting that Meta also uses the Hugging Face hosting platformopenMobileLLM-R1 is released under the Apache 2.0 license, making it easy for developers to download and use. It can be directly run with the vLLM inference engine. Deployment is accomplished by simply registering the model architecture to the ModelRegistry as Llama4ForCausalLM. For developers seeking this, this means building dedicated AI applications on mobile devices at a lower cost, without relying entirely on cloud resources.
Overall, the MobileLLM-R1 represents another step forward for Meta's "small but precise" AI model strategy. By focusing on inference capabilities and reducing resource requirements, it enables AI to be truly integrated with users' devices and daily lives. As more manufacturers advance on-device AI solutions, the AI inference capabilities of smartphones, laptops, and even IoT devices will usher in a new wave of upgrades.
