Cohere Opens 2.4B Vision Model for Edge Deployment
North Micro Vision packages document and visual understanding into a 2.4B model with downloadable weights and an Apache 2.0 license.
A compact, customizable vision model
Cohere has released North Micro Vision, a 2.4-billion-parameter vision-language model intended for edge deployment and specialized visual applications. The weights are available from Hugging Face under the Apache 2.0 license, permitting commercial use, modification and redistribution without dependence on Cohere’s hosted service.
The model is designed around structured information extraction, visual question answering and the interpretation of documents, charts and scientific figures. Cohere’s published comparisons place it ahead of similarly compact models including Gemma 4 E2B and Ministral 3 3B on a selection of vision benchmarks. As with any vendor-selected evaluation, those results need independent validation on the images, languages and document layouts encountered in production.
North Micro Vision’s size is central to the release. At 2.4 billion parameters, it is small enough to target Apple devices and modest GPU installations rather than only data-center clusters. Community support was available at launch for Apple’s MLX framework, while Axolotl and Nvidia tooling provide paths for fine-tuning and deployment. Cohere is explicitly presenting the base model as something organizations can turn into narrow visual experts—for example, extracting fields from a particular form or interpreting a specialized category of technical diagram.
Why it matters
Many high-value vision tasks do not require a general assistant capable of answering every question. They require a controllable model that processes sensitive documents cheaply, runs near the data and can be adapted to a fixed schema. A permissively licensed model in this size class can therefore matter more to enterprise deployment than a modest benchmark gain by a much larger system. North Micro Vision also strengthens a broader open-model trend: capabilities once reserved for cloud multimodal APIs are being separated into compact components that device makers and software vendors can own, tune and operate themselves.