MobiLlama is a small language model from MBZUAI’s Oryx research group, tailored specifically for edge and mobile devices. At around half a billion parameters, it aims to deliver useful language abilities within the tight memory and compute budgets of phones and embedded hardware. The work was spotlighted at ICLR 2025, and the repository has about 671 stars on GitHub.
Most open LLMs assume a server GPU. MobiLlama starts from the opposite constraint: what is the best model you can actually run on-device? Its design emphasizes parameter efficiency, using a carefully shared architecture to keep the footprint small while retaining competitive performance for its size class on common benchmarks. Training data and evaluation details are documented in the repository and paper, which appeared in the ICLR 2025 small language model spotlight track. Released under the Apache-2.0 license, the weights are available for researchers and product teams building offline, private or low-latency language features into mobile and edge products.
Its small footprint also makes it a convenient testbed for quantization, distillation and on-device inference research, where fast iteration matters more than raw capability, and a useful reference when sizing models for real hardware limits.
MobiLlama is free and open source under the Apache-2.0 license. There are no API fees or subscriptions; you download the weights and run them on your own devices.
MobiLlama is for mobile and embedded developers who need language features that work offline and respect privacy, and for researchers studying small language models. It is a practical baseline when a cloud LLM is too big, too slow or too costly per query.