On August 12, Cohere Labs released North-Micro-Vision-Instruct on Hugging Face: a 2.4B-parameter open-weight vision-language model (VLM) under the Apache 2.0 license. It is Cohere's smallest VLM to date, and its positioning is explicit — not an all-rounder, but a compact, fine-tunable foundation for vertical scenarios like document understanding, OCR, and visual grounding.

Native Resolution Is the Headline Feature

Most VLMs compress every input into a small fixed square image as their first step; aspect ratios, small fonts, and table lines are lost in that squeeze. North Micro Vision goes the other way: visual inputs stay at native resolution, capped at 1654×2339 pixels — exactly an A4 page scanned at 200 dpi — while preserving the aspect ratio. For inputs like receipts, forms, and financial-report screenshots where "the detail IS the information", this design matters more than raw parameter count.

Architecture: Three Parts, Each with Pedigree

The model has three components:

  • Vision encoder: 400M parameters, custom-trained, continued from Google's SigLIP 2 SO400M checkpoint. The key technique is C-RoPE — combining 2D RoPE with bilinearly interpolated 1D positional embeddings to support native-resolution inputs.
  • Projector: follows the DeepStack approach, injecting patch embeddings from multiple vision-encoder layers into early LLM layers, giving the language model access to visual representations at multiple levels of abstraction.
  • Language model: the 2B-parameter North Micro LLM, inheriting the Command A+ hybrid attention architecture — three sliding-window attention layers with RoPE interleaved with one global attention layer without positional embeddings.

Training: A "Coarse-to-Fine" Resolution Ramp

Training ran in four stages: first aligning the vision encoder and projector on 10M examples at fixed 384×384 resolution (60% dense captions + 40% OCR); then ramping resolution to 1024×1024 and A4@200dpi in two steps with the full model trained jointly; a third stage of multimodal instruction tuning on 50M examples, where OCR, charts/tables, and grounding/counting together account for over 60% of the mix; and finally 500k preference examples for safety and formatting — using a simplified MPO variant (dropping the BCO loss, combining DPO with a 15%-weighted SFT auxiliary loss).

This OCR-heavy curriculum shows up directly in the scorecard.

Benchmarks: Genuinely Strong Where It Counts, Weak Where It Doesn't

The comparison group includes Ministral-3 (3.8B), LFM2.5-VL (1.6B), Phi-3.5-vision (4.2B), Gemma-4-E2B (5.1B), Qwen3-VL-2B, Qwen3.5-2B, and SmolVLM 2.2B:

Benchmark North Micro Vision Notes
DocVQA 0.921 Second only to Qwen3.5-2B's 0.926
ChartQA 0.808 Best in the comparison group
OCRBench 0.792 Behind only LFM2.5-VL and Qwen3.5
RefCOCO (avg) 0.732 vs SmolVLM 2.2B at 0.018, Ministral-3 at 0.317
MMMU 0.329 Bottom of the group
MMLU 0.504 Clearly below general-purpose models

The most striking number is RefCOCO: 0.732 vs 0.018 — visual grounding capability in same-size open models spans orders of magnitude. Bounding boxes were normalized to a 0–1000 coordinate scale during training, and grounding/counting data took 13.3% of instruction tuning. Conversely, MMMU at 0.329 and MMLU at 0.504 mean general knowledge and STEM reasoning were deliberately given up. Cohere's own conclusion: "benchmark profile strongest on document understanding and visual grounding."

Niche and Commentary

Cohere's home turf has been enterprise text models (the Command R and Aya lines); this release is its step into open-source multimodal under a fully permissive license. The official blog frames it under "sovereign AI": clear licensing, open weights, transparent evaluation. Ecosystem support came fast: MLX-VLM weights contributed by the community (runs on Apple Silicon), an NVIDIA AutoModel fine-tuning recipe, Axolotl support, with official vLLM support "coming soon".

My take: the 2-3B "specialist model" is becoming its own track. LFM2.5-VL-3B targets phones, Muse Glimmer 30B targets local agents, and North Micro Vision stakes its claim on "native resolution + documents/grounding". The lesson for model selection is direct — aggregate leaderboard scores are useless; look at your task profile. If your workload is invoice recognition, form extraction, or screenshot QA, a 2.4B specialist may beat a 5B generalist while costing less; asking it to do math or open-domain QA is simply using the wrong tool.

The open license leaves room for adaptation, and documents are exactly the domain enterprises are most willing to fine-tune for. This "small and sharp" playbook deserves watching.

Reference: https://huggingface.co/blog/CohereLabs/meet-north-micro-vision-instruct