2.4B total (2B language model + 400M vision encoder)
128K text backbone; multimodal training and validated operation up to 8K
Apache 2.0
Compact open-weight vision-language model built for prototyping, task-specific fine-tuning, and specialized multimodal applications. North Micro Vision uses Cohere's CohereCompass architecture: a 28-layer, 2B language backbone interleaving 4K sliding-window attention and full attention in a 3:1 schedule, paired with a custom-trained 400M vision encoder initialized from SigLIP 2 SO400M. It processes native-resolution images without forcing a fixed aspect ratio, accepts interleaved and multiple images, and supports visual question answering, captioning, OCR, document and chart understanding, multilingual image understanding, object counting, and grounding with normalized 0-1000 bounding boxes. It has a 262,144-token vocabulary and supports English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese, Arabic, and additional languages. Official evaluations report 0.921 on DocVQA, 0.808 on ChartQA, 0.792 on OCRBench, 0.775 on AI2D, 0.732 average on RefCOCO, 0.725 on CountBench, and 0.622 on RealWorldQA. The release is distributed as a single approximately 5 GB BF16 safetensors checkpoint with Transformers support. It is not presented as a reasoning, coding, tool-calling, or agentic model; system prompts are not recommended because they were not used during training, and multimodal contexts beyond 8K rely on unbenchmarked extrapolation.