North Micro Vision Instruct

// Overview

North Micro Vision Instruct is a small vision-language model published by Cohere Labs on 12 August 2026, extending the North family — until then a text and coding line — to images. At 2.4 billion parameters it is aimed squarely at prototyping, task-specific fine-tuning and specialised multimodal applications rather than at leaderboard positions: Cohere Labs describes it as a compact foundation to build on, and ships it under Apache 2.0 with no per-token API price.

The model combines a 400M vision encoder, custom-trained from SigLIP 2 SO400M and using 2D RoPE alongside learned 1D positional embeddings, with a 2B in-house language model that interleaves sliding-window and global attention. Rather than passing a single pooled image representation into the language model, it injects patch embeddings taken from several encoder layers into the language model's early layers. Images are processed at their native resolution with the aspect ratio preserved, and multiple images can be supplied in one conversation. The language model's context runs to 128K tokens, though Cohere Labs validates multimodal inputs only to 8K.

Eleven languages are declared on the model card — English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese and Arabic. Its strongest published results are on document and chart understanding (0.921 on DocVQA and 0.808 on ChartQA) and on grounding and counting (0.732 on RefCOCO-average and 0.725 on CountBench), where the native-resolution design pays off most directly.

Cohere Labs' launch table sets it against seven other small vision-language models. The picture is mixed by design: on broad VQA rows such as MMBench (0.687) and MMStar (0.518) it sits mid-pack behind Qwen3.5-2B and Qwen3-VL-2B, while on RefCOCO it leads every listed model except Qwen3.5-2B by a wide margin — 0.732 against 0.317 for Ministral-3-3B and 0.084 for Gemma-4-E2B. A LiteRT conversion of the checkpoint is published by the litert-community organisation for on-device use.

// Benchmarks

This model's scores

1. DocVQA (val)0.921
2. ChartQA (test)0.808
3. OCRBench0.792
4. AI2D (test)0.775
5. RefCOCO (average)0.732
6. CountBench0.725
7. MMBench DEV EN v1.10.687
8. InfoVQA (val)0.652
9. RealWorldQA0.622
10. HallusionBench0.615
11. GQA (test-dev balanced)0.574
12. MMStar0.518
13. OCRBench v2 (en)0.367

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Apache-2.0 open weights, self-hostable with no per-token pricing
• Native-resolution image processing that preserves aspect ratio, with multi-image inputs in one conversation
• 0.921 on DocVQA and 0.808 on ChartQA — document and chart reading is the model's strongest published area
• 0.732 on RefCOCO-average and 0.725 on CountBench, far ahead of most models in its size class on grounding and counting
• 2.4B parameters total, small enough for laptop-class hardware, with a LiteRT build published for on-device runtimes

// Best for

• Document, form, chart and receipt understanding where the input is a scan or screenshot rather than clean text
• Task-specific fine-tuning — Cohere Labs positions the model as a compact base to specialise rather than a finished product
• Visual grounding and object counting in applications that need boxes or tallies, not just captions
• On-device and edge multimodal features, using the LiteRT conversion of the checkpoint