Cohere Labs Ships North Micro Vision to Read Full Documents on One GPU

Cohere Labs has quietly become one of the more interesting shops in the open-weight VLM space, and its latest release fits that pattern. North Micro Vision Instruct is a 2.4B-parameter open-weight vision-language model with native-resolution image support, released under Apache 2.0. It is the smallest VLM Cohere has shipped so far, and the pitch is straightforward: small enough to fine-tune on your own hardware, capable enough to read a full A4 page without downsampling it into mush.

A tiny model built for documents

The design choice that shapes everything else is native-resolution input. Most compact VLMs resize every image to a fixed square (often 384x384 or 448x448), which is fine for cat photos and useless for a spreadsheet. North Micro Vision preserves full document, chart, and screenshot detail up to A4 at 200 dpi. That single decision is why it punches above its weight on document-heavy tasks.

Cohere is explicit about the intended shape of the model. It is meant for:

• Prototyping and task-specific fine-tuning
• General visual question answering and image captioning
• Multilingual and multi-image understanding
• Visual grounding and spatial reasoning
• OCR, chart and document understanding, and structured information extraction

It supports 11 languages including English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese and Arabic, and the tokenizer has a 262,144-token vocabulary. The LM backbone technically has a 128K context window, though only the 8K-token multimodal range is validated.

How the architecture is stitched together

The model has three parts. North Micro Vision combines a custom-trained 400M-parameter native-resolution vision encoder with an in-house 2B-parameter language model called North Micro LLM. The language model reuses Cohere's Command A+ recipe: three sliding-window attention layers with rotary positional embeddings interleaved with one global attention layer that has no positional embeddings. NoPE on the global layer is the trick that lets the model handle long contexts without positional drift.