Cohere Releases 2.4B Vision Mini Model: Document Understanding Surpasses Ministeer AI 3.3B

According to DeepBeat Monitoring, Cohere has open-sourced its smallest visual language model to date, North Micro Vision, with only 2.4B parameters, released under the Apache 2.0 license. It focuses on documents, tables, charts, screenshots, and OCR, capable of processing images in their original aspect ratio and resolution without the need for prior resizing.

The model consists of a 2B language model and a 400M visual encoder. In official testing, DocVQA achieved a score of 92.1%, surpassing Ministral 3 3B at 89.6% and Gemma 4 E2B at 73.2%, coming close to Qwen3.5-2B at 92.6%. The visual localization task RefCOCO scored 73.2%, also significantly higher than Ministral and Gemma.

However, it is not the strongest model in terms of size. Qwen3.5-2B still outperforms it on most general vision, OCR, and multimodal benchmarks, while the strength of North Micro Vision lies more in document understanding and visual localization. It is also not an inference model, does not support tool invocation, and has limited mathematical and coding capabilities.