North Micro Vision Packs Document AI Into 2.4B Params

Most vision-language models solve the resolution problem by ignoring it: every image gets squashed to a small fixed square, and whatever detail lived in the fine print goes with it. Cohere Labs took the opposite approach with North-Micro-Vision-Instruct, published on August 12, 2026 — a 2.4-billion-parameter model that reads images at their native resolution and is small enough to run on a laptop.

• 2.4B parameters total: a 400M vision encoder paired with a 2B language model, released under Apache 2.0
• Processes images at native resolution up to roughly 1654x2339 pixels, the size of an A4 page scanned at 200 dpi
• Reports 0.921 on DocVQA, 0.808 on ChartQA, and 0.732 on RefCOCO visual grounding
• Targeted at local, edge, and specialized deployments, with MLX quantizations and NVIDIA AutoModel recipes available

Why Native Resolution Changes Document AI

The standard pipeline resizes an input image to a uniform square — often 336 or 448 pixels on a side — before the vision encoder ever sees it. That is fine for identifying a dog. It is fatal for a scanned invoice, where the line items, the footnotes, and the handwritten annotation in the margin are precisely the pixels being thrown away.

North Micro Vision preserves aspect ratio and fine detail instead, handling a full-page scan without downsampling it into illegibility. That single design decision is what the benchmark profile reflects: the model's strongest results cluster in DocVQA, ChartQA, OCR, and visual grounding — the tasks where detail preservation is the whole game — rather than in general image captioning, where a smaller model would be outgunned.

What Can a 2.4B Vision Model Actually Do?

Cohere Labs positions it for dense visual understanding: documents, tables, charts, and screenshots, with multilingual coverage across visual domains, plus grounding and counting. The reported 0.921 on DocVQA is the headline figure, and 0.808 on ChartQA suggests it holds up when the information is encoded as a graph rather than as text.

What it is not is a general-purpose frontier multimodal model. At 2.4B parameters it will not reason its way through a hard research problem the way a much larger open-weight release does. The trade is deliberate — a specialist that fits where a generalist cannot.

Running North Micro Vision on Local Hardware

The deployment story is the reason to care. At 2.4B parameters and quantized, this comfortably fits laptop- and mobile-class hardware, which puts full-page document understanding inside the same envelope as an edge box or a mini PC on a shelf. Ecosystem support at launch includes MLX quantizations for Apple silicon, NVIDIA AutoModel recipes, and Axolotl for fine-tuning; vLLM support was listed as coming soon.

Apache 2.0 licensing removes the other obstacle. Document processing is exactly the workload organizations are least willing to send to a hosted API — contracts, medical records, financial statements — and a permissive license on a model that runs in-house sidesteps the entire question. It fits the same on-device trend as Liquid AI's LFM2.5-VL-3B, where the achievement is not raw capability but capability that fits in three gigabytes.

The Case for Small Specialist Models

There is a quiet argument running underneath releases like this one, and it is worth stating plainly: for a large share of real production work, the right model is not the biggest one. It is the smallest one that clears the bar, running close to the data, under a license that does not complicate the deployment.

A 2.4B model that reads a full-resolution page well beats a frontier model you cannot legally point at the documents in question. Weights are on Hugging Face now, and the broader open-weight picture is something we track across our AI coverage.

Sources: Cohere Labs on Hugging Face — August 12, 2026; North-Micro-Vision-Instruct model card — accessed August 17, 2026.