Namo-R1 is an open-source small multimodal large language model (MLLM) project focused on achieving usable vision-language understanding capability with an extremely small parameter count. The first model series, Namo-500M, has only 500 million parameters and can run directly on a CPU. The project team claims that at the same scale it surpasses SmolVLM and Moondream2.
In terms of use cases, Namo-R1 is aimed at those who want to run vision-language models on edge devices or in environments without GPUs, as well as researchers and developers who want to train their own MLLM from scratch. The project emphasizes being fully open source, including the training data curation pipeline, model architecture, and training scheduler, in order to reproduce AI research.
Main capabilities include: CPU-friendly, and able to run fairly quickly on a CPU; support for multilingual OCR, covering Chinese, English, Japanese, and other languages; general concept understanding; image captioning; native dynamic resolution input, able to adapt to images of any aspect ratio; support for R1 for post-training; and an extensible design for future audio (ASR/TTS) and cross-modal fusion.
In terms of usage, the project provides a simple demo. You can run python demo.py in the terminal for CLI multi-turn chat, and afterward you can obtain an OpenAI-like service interface locally. The command-line direct run Namo cli feature is said to be provided later. For training, the project has open-sourced all model code, training steps, and scripts, and supports training with SigLIP2 as the vision encoder.
Applicable scenarios include: running vision-language models in environments without GPUs, edge deployment, multilingual OCR, image understanding and captioning, and research work on training or post-training small MLLMs from scratch.
In terms of limitations, the project team explicitly states that current testing has only been conducted on a limited number of benchmarks, and more metrics will be reported in the future; larger-scale models (up to 7B) and more vision encoders are still in the training plan. When using deepspeed, if you encounter the error "no_sync context manager is incompatible with gradient partitioning logic of ZeRO stage 2", it is recommended to upgrade transformers to 4.48+ and use the latest version of deepspeed. The code is released under the MIT license, and the copyright belongs to the Namo authors. If you need to join the community, you can visit its Discord:https://discord.gg/5ftPBVspXj 。
