이미지: METAL LAB 생성
Summary
- Cohere has released North Micro Vision, an ultra-compact vision-language model with roughly 2.4 billion parameters, distributing it on Hugging Face under an Apache 2.0 license
- On document and chart understanding benchmarks such as DocVQA and ChartQA, it outperformed Gemma 4 E2B and Ministral 3 3B in most cases, but trailed slightly on multilingual understanding (MMMB)
- The model supports native-resolution processing, multi-turn conversation, visual grounding, and multilingual image understanding, and is designed for experimenting with document automation
- 모델명
- North Micro Vision (North-Micro-Vision-Instruct)
- 개발사
- Cohere
- 파라미터 규모
- 24억(2.4B)
- 라이선스
- Apache 2.0
- 공개처
- Hugging Face
- 공개일
- 2026-08-12
- DocVQA 점수
- 0.921 (Ministral 3 3B 0.896, Gemma 4 E2B 0.732)
- RefCOCO 점수
- 0.732 (Ministral 3 3B 0.317, Gemma 4 E2B 0.084)
Eyes for reading documents, now smaller
Getting a computer to properly read a single scanned contract used to require a bigger model than one might expect. North Micro Vision, unveiled by Cohere on August 12, arrives at a size that upends that assumption. Its parameter count is 2.4 billion (2.4B) — putting it among the smallest of recent open-source vision-language models (VLMs, models that understand images and text together). Cohere introduced it as the newest member of its "North" model family, describing it as "the smallest vision-language model."
The reason small size matters is deployment cost. VLMs typically stack an image encoder on top of a language model, so they consume more memory than a text-only model with the same parameter count. At the 2.4B scale, real-time inference becomes feasible even on laptops or small servers, dramatically lowering the barrier for developers who want to experiment with document automation.
What it can do
Cohere lists four capabilities. Native-resolution processing, which handles images while preserving their original resolution; multi-turn conversation about an image; spatial reasoning and visual grounding, which pinpoint specific regions within an image; and multilingual image understanding. Cohere emphasized that the model is "ideal for document understanding," and the published benchmark results indeed cluster around that strength.
Where it stands on benchmarks
The table Cohere released places North Micro Vision alongside Ministral 3 3B, LFM2.5-VL-1.6B, Phi-3.5-vision-instruct, Gemma 4 E2B, and Qwen3.5-2B for score comparison. Gemma 4 comes from Google DeepMind, and the Qwen3 series comes from Alibaba Qwen — both names already registered in our ontology.
| Benchmark | North Micro Vision | Ministral 3 3B | Gemma 4 E2B | Qwen3.5-2B |
|---|---|---|---|---|
| DocVQA (document understanding) | 0.921 bar:92 | 0.896 bar:90 | 0.732 bar:73 | 0.926 bar:93 |
| ChartQA (chart understanding) | 0.808 bar:81 | 0.791 bar:79 | 0.422 bar:42 | 0.775 bar:78 |
| RefCOCO (visual grounding) | 0.732 bar:73 | 0.317 bar:32 | 0.084 bar:8 | 0.785 bar:79 |
| MMMB (multilingual understanding) | 0.728 bar:73 | 0.734 bar:73 | 0.743 bar:74 | 0.745 bar:75 |
The table largely backs up Cohere's claims. North Micro Vision beat Ministral 3 3B and Gemma 4 E2B — both larger or comparably sized — on nearly every document, chart, and grounding metric. On RefCOCO in particular — a benchmark measuring the ability to pinpoint a specific object in an image from a text instruction — it left Gemma 4 E2B (0.084) far behind. However, on MMMB, which measures multilingual image understanding, it trailed slightly behind Ministral 3 3B (0.734) and Gemma 4 E2B (0.743), and on document understanding it also fell short of the smaller Qwen3.5-2B (0.926). Against the two models Cohere explicitly named as comparisons, its lead holds up, but once the other small models in the table are factored in, the results shift back and forth by category.
Why open-source
Cohere CEO Aidan Gomez recently named three conditions an open-source model needs: customizability, cost-efficiency, and accessibility. Releasing North Micro Vision under Apache 2.0 — a license with few restrictions on commercial use and redistribution — falls in line with that stance. Cohere has posted the weights on Hugging Face so anyone can download and experiment with them.
Open-sourcing small VLMs has been a trend among competitors as well, with companies like Moonshot AI releasing similar models over the past month. However, North Micro Vision is far smaller in parameter count than those, positioning itself around deployment convenience rather than raw performance.
So what changes
Automatically reading scanned documents and converting them into data — tasks like sorting receipts, reviewing contracts, or extracting numbers from charts — has often relied on large cloud APIs. With a 2.4-billion-parameter model like North Micro Vision scoring near the top on document-understanding benchmarks, there's now one more option for running such tasks locally and for free. Its performance isn't always the best, but its results relative to its size could serve as a practical reference point for developers weighing real-world deployment.
