Cohere's North Micro Vision Reads Full A4 Documents Without Losing Detail

Cohere just dropped North Micro Vision Instruct, a 2.4B-parameter vision-language model (VLM) built specifically for document-heavy workloads. It is open-weight, free to use commercially, and available right now under an Apache 2.0 license. The headline feature is native-resolution image processing: instead of squashing every image into a fixed square before feeding it to the model, North Micro Vision processes images at their actual dimensions and aspect ratios.

Why native resolution matters

Most small VLMs resize inputs to a fixed resolution (say, 224x224 or 336x336 pixels) before encoding them. That is fast, but it destroys fine-grained detail: small text in a scanned PDF becomes illegible, table borders blur together, and chart labels vanish. North Micro Vision preserves the aspect ratio and fine detail of documents, tables, charts, screenshots, and forms instead of first reducing every input to a small square image.

The model supports native-resolution inputs up to 1654 x 2339 pixels, corresponding to an A4 page at 200 dpi. That is enough resolution to read dense legal text, financial tables, or handwritten forms without any preprocessing tricks on your end.

Under the hood: a three-part architecture

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 follows Cohere's Command A+ architecture, interleaving three sliding-window attention layers that use rotary positional embeddings with one global attention layer without positional embeddings. Sliding-window attention (SWA) lets each token attend only to a local neighborhood, keeping memory costs low on long sequences, while the periodic global layer lets information flow across the full context.

The glue between vision and language is a projector that maps visual features into the language model's token space. Following the DeepStack approach, patch embeddings from multiple vision-encoder layers are injected into corresponding early LLM layers, giving the language model access to visual representations at different levels of abstraction. Think of it as giving the language model a multi-scale view of the image rather than a single flattened summary.

How it was trained

Training proceeded in four stages: Stage 1 adapted the vision encoder and projector; Stages 2.1 and 2.2 increased resolution while jointly training the encoder, projector, and language model; Stage 3 instruction-tuned the full model; and Stage 4 used a simplified variant of Mixed Preference Optimization (MPO) to improve safety, formatting, and response quality. MPO is a technique that trains a model to prefer good responses over bad ones using human or AI-labeled preference pairs, similar to RLHF but without a separate reward model.

The vision encoder was continued from the SigLIP 2 SO400M checkpoint, with a curriculum that progressively increased image resolution and incorporated Continual Rotary Position Embedding (C-RoPE) to support native-resolution inputs. The instruction-tuning data mix was deliberately balanced across domains:

• Native OCR: 17.8%
• Charts and tables: 17.8%
• Grounding and counting: 13.3%
• OCR QA: 13.3%
• General VQA: 11.2%
• Captioning and knowledge: 8.9%
• Text-only: 8.9%
• Mathematics and Science: 8.8%

The training data drew on publicly available datasets and an in-house, large-scale multilingual document corpus that supported the synthesis and curation of data for OCR, document understanding, chart understanding, captioning, HTML table generation, and visual grounding across languages.

Where it shines and where it doesn't

The benchmark story is strongest where the training data was heaviest. North Micro Vision scores 0.921 on DocVQA and 0.808 on ChartQA, outperforming all compared models except Qwen3.5-2B on DocVQA. On RefCOCO visual grounding, it scores 0.732 averaged across all splits, far ahead of Ministral-3-3B (0.317), Gemma-4-E2B (0.084), and SmolVLM2.2B (0.018).

The weak spots are equally clear. The model is not a reasoning model and has limited math and code-generation capabilities. On MMMU (a graduate-level STEM benchmark), it scores 0.329, the lowest of all compared models. Tool calling and agentic workflows are not supported, and system prompts are not recommended because the model was not trained with them. If you need a general-purpose chat assistant or a model that can write code, this is not the right tool.

There is also a memory caveat worth flagging: native-resolution inputs can increase memory use and latency as image dimensions grow. Processing a full A4 page at 200 dpi will cost more VRAM than a thumbnail, so plan your hardware accordingly.

The practical sweet spot

The model's design makes it a natural fit for a specific class of problems:

• Document intelligence pipelines -- invoice extraction, contract review, form parsing
• Chart and table understanding -- reading financial reports, dashboards, or research figures
• Multilingual OCR -- supported languages include English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese, and Arabic
• Visual grounding -- returning bounding boxes for objects in an image, normalized to a 0-1000 scale
• Fine-tuning base -- at 2.4B parameters, it is small enough to fine-tune on a single GPU

With the right inference stack and quantization, models at this scale can support experimentation beyond server-only deployments, including on laptops and edge or mobile-class hardware. Quantized versions are already appearing on Hugging Face and can be run via llama.cpp, Ollama, or LM Studio.

Getting started

The model requires Transformers 5.16.0, which is not yet on PyPI. Install from source for now:

uv pip install accelerate pillow
uv pip install "git+https://github.com/huggingface/transformers.git"

A minimal inference example looks like this:

from transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "CohereLabs/North-Micro-Vision-Instruct"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, dtype="auto", device_map="auto"
)
messages = [{
    "role": "user",
    "content": [
        {"type": "image", "url": "https://your-image-url.com/doc.png"},
        {"type": "text",  "text": "Extract all line items from this invoice."},
    ],
}]
inputs = processor.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True,
    return_tensors="pt", return_dict=True
).to(model.device)
outputs = model.generate(
    **inputs, max_new_tokens=512,
    do_sample=True, temperature=0.7, top_p=0.8, top_k=20
)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Public vLLM support is coming soon. Until then, Transformers is the recommended path. SGLang is also supported today via the OpenAI-compatible API if you need a server setup.

The bigger picture

North Micro Vision lands in a crowded sub-3B VLM space alongside SmolVLM2, LFM2.5-VL, and Qwen3-VL-2B. What differentiates it is the deliberate focus on document fidelity over raw benchmark averages. The release reflects Cohere's broader work on sovereign AI by pairing model development with clear licensing, open weights, and transparent evaluation. For teams building document pipelines that need to run on-prem, avoid API costs, or fine-tune on proprietary data, a permissively licensed 2.4B model with genuine document understanding is a meaningful addition to the toolkit.