Ddos 
As generative AI continues to evolve toward lighter and more localized deployments, Meta has announced the open-sourcing of its latest MobileLLM-R1 series models, designed to run directly on mobile devices and optimized for reasoning in mathematics, programming, and scientific domains. This move not only responds to the growing demand for on-device AI but also underscores the company’s ongoing commitment to advancing reasoning capabilities.
MobileLLM-R1 is the newest addition to the Meta MobileLLM family, distinguished by its dual focus on efficiency and specialization. The series includes both base models and supervised fine-tuned (SFT) variants, with parameter sizes of 140M, 360M, and 950M. The base models support a context length of 4K tokens, while the fine-tuned versions extend this to 32K tokens, significantly enhancing their ability to handle complex tasks.
Meta emphasizes that MobileLLM-R1 is not a general-purpose chatbot model but one specifically engineered for targeted reasoning use cases. These include solving mathematical problems, writing code (across languages such as Python and C++), and tackling tasks related to scientific research.
Despite the fact that the largest model, MobileLLM-R1 950M, was trained on less than 5TB of high-quality data (with only 2TB of pretraining tokens), its performance is remarkable. According to Meta, MobileLLM-R1 outperformed Qwen 3 0.6B — which was trained on 36TB of data — across benchmarks such as MATH, GSM8K, MMLU, and LiveCodeBench.
In finer-grained comparisons, MobileLLM-R1 950M achieved accuracy on the MATH benchmark five times higher than Olmo 1.24B and twice that of SmolLM 1.7B, while also leading decisively in code generation and problem-solving. Even the smallest variant, MobileLLM-R1 140M (base), outperformed SmolLM2-135M, while the 360M version surpassed both Gemma-3-270M-pt and SmolLM2-360M (base) by a wide margin — highlighting Meta’s architectural and training optimizations.
Notably, Meta has also released MobileLLM-R1 on the Hugging Face platform under the Apache 2.0 license, enabling developers to freely download and deploy the models. They can be run directly with the vLLM inference engine by registering the architecture as Llama4ForCausalLM in the ModelRegistry. For developers, this means building specialized AI applications on mobile devices at lower costs, without being entirely dependent on cloud resources.
Taken as a whole, MobileLLM-R1 marks another significant milestone in Meta’s “small yet powerful” AI strategy. By focusing on reasoning ability while reducing resource requirements, the models bring AI closer to users’ personal devices and everyday contexts. As more companies advance edge AI solutions, the reasoning power of smartphones, laptops, and even IoT devices is poised for a new wave of transformation.
Related Posts:
Support Our Threat Intelligence
If you find our CVE report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
