Cohere releases tiny vision model with native resolution and open weights

Cohere has released North-Micro-Vision-Instruct, a vision-language model (VLM) of 2.4 billion parameters under an Apache 2.0 license that supports the native resolution of input images, avoiding compression of inputs to a small square and loss of detail in documents, tables or screenshots. The model surpasses Gemma 4 E2B and Ministral 3 3B on general vision understanding benchmarks and records especially strong results in document understanding and visual Q&A.

The architecture consists of three components: a 400 million-parameter vision encoder adapted for native resolution, a projector that maps visual features into the language model’s embedding space, and a compact 2 billion-parameter language model called North Micro LLM. The language model follows Cohere’s Command A+ architecture, employing three sliding-window attention layers with rotary positional embeddings (RoPE) and a single global attention layer without positional embeddings. The encoder combines 2D RoPE with learned 1D positional embeddings to preserve spatial structure in full-resolution inputs.

The projector implements the DeepStack method: patch embeddings from multiple vision-encoder layers are injected into early parallel layers of the language model, providing visual representations at varying abstraction levels early in processing. This architectural choice is intended to improve fine-grained understanding of details in documents and diagrams without dramatically increasing parameter count.

Training proceeded in four stages, with stage 2 split into two resolution phases. Stage 1 aligned the encoder and projector; stages 2.1 and 2.2 increased resolution and context length while jointly training all components; stage 3 performed instruction tuning on the full model; and stage 4 applied a simplified Mixed Preference Optimization (MPO) to enhance safety, response style and overall quality. The vision encoder continued pre-training from Google’s SigLIP 2 SO400M checkpoint, using a curriculum that gradually raised resolution and context length and incorporated C-RoPE, a blend of 2D RoPE with linearly mixed 1D embeddings.

At 2.4 billion parameters the model is a natural candidate for fine-tuning on domain-specific data and for deployment on local hardware, laptops, edge devices and mobile, provided an appropriate inference stack and quantization are used. Cohere notes that a public vLLM will be available soon. The weights are already posted on Hugging Face under CohereLabs/North-Micro-Vision-Instruct.

The release is part of Cohere’s broader sovereign AI effort, emphasizing clear licensing, open weights and transparent evaluation. Unlike closed models or releases lacking benchmark methodology, Cohere publishes detailed test results against direct competitors in the compact-model category. For developers building multimodal applications in Hebrew or Arabic who require full model control without reliance on external APIs, the model offers a tool worth close examination, especially when processing complex documents on local hardware.