North Micro Vision Instruct

North Micro Vision Instruct

Quick answer: North Micro Vision Instruct is Cohere Labs' 2.4B-parameter open-weight vision-language model, released August 12, 2026 under the Apache 2.0 license. It pairs a custom 400M-parameter native-resolution vision encoder (built from SigLIP2 SO400M) with an in-house 2B-parameter "North Micro LLM" following Cohere's Command A+ architecture. It leads its size class on document understanding (92.1% DocVQA, 80.8% ChartQA, 79.2% OCRBench) and grounding/counting (73.2% RefCOCO-avg, 72.5% CountBench), while trailing on knowledge-heavy STEM reasoning (32.9% MMMU) and instruction-following (30.7% MMLU-Pro). Weights are on Hugging Face; native-resolution input means it preserves full document/chart/screenshot detail up to A4-page-at-200-dpi rather than downsizing to a small square image.

At a Glance

Where North Micro Vision Instruct leads

• 92.1% DocVQA (val) — the strongest document-understanding score in its launch comparison set, ahead of Ministral-3-3B-Instruct, LFM2.5-VL-1.6B, Phi-3.5-vision-instruct, gemma-4-E2B-it, and Qwen3.5-2B
• 73.2% RefCOCO-avg grounding — by far the strongest in the comparison set, well ahead of the next-best LFM2.5-VL-1.6B (58.1%)
• 79.2% OCRBench and 77.5% AI2D — consistently strong text-recognition and diagram-reading scores, with a competitive 80.8% ChartQA (test), just behind Phi-3.5-vision-instruct's 82.1%
• Native-resolution vision encoder preserves aspect ratio and fine detail for documents, tables, charts, screenshots, and forms — up to a full A4 page at 200 dpi — instead of first downsizing every input to a small square
• 2.4B total parameters (400M vision encoder + 2B language model), practical to fine-tune for specialized visual domains and deployable on laptops or edge/mobile-class hardware with the right inference stack

Where it lags

• 32.9% MMMU (dev/val) — knowledge-heavy STEM reasoning trails larger comparison models like Ministral-3-3B-Instruct (50.8%)
• 30.7% MMLU-Pro and 37.3% Multi-IF — text-only knowledge and complex instruction-following are not this model's strongest areas relative to larger rivals
• 36.7% OCRBench v2 (English) — the harder, localization- and reasoning-heavy OCR benchmark remains difficult across the entire comparison set, not just this model

Best for: compact, edge-aware document and chart understanding, multilingual visual QA, and visual grounding/counting — workloads where preserving native document resolution and fine-tuning for a specific visual domain matter more than frontier-scale general knowledge.

What North Micro Vision Instruct Is

North Micro Vision Instruct is Cohere Labs' smallest vision-language model to date, designed as a compact foundation for specialized multimodal applications rather than a general frontier competitor. Its three-part architecture pairs a custom-trained 400M-parameter native-resolution vision encoder — continued-pretrained from the google/siglip2-so400m-patch16-384 checkpoint with Continual Rotary Position Embedding (C-RoPE) to support variable, native-resolution inputs — with a DeepStack-style projector that injects patch embeddings from multiple vision-encoder layers into corresponding early layers of the language model, giving it visual representations at several levels of abstraction rather than a single flattened embedding.

The language model, "North Micro LLM," is a 2B-parameter in-house model following Cohere's Command A+ architecture, interleaving three sliding-window attention layers using rotary positional embeddings with one global attention layer that has no positional embeddings. Training proceeded in four stages: Stage 1 adapted the vision encoder and projector at a fixed 384x384 resolution with the language model frozen; Stages 2.1 and 2.2 progressively raised resolution (to 1024x1024, then to 1654x2339 — a full A4 page at 200 dpi) while jointly training the encoder, projector, and language model; Stage 3 instruction-tuned the full model on a diverse multimodal mixture; and Stage 4 applied a simplified variant of Mixed Preference Optimization (MPO) for safety, formatting, and response quality.

On Benchgen's read, North Micro Vision Instruct's standout results cluster tightly around document, chart, and OCR understanding — exactly the native-resolution, layout-preserving use case its architecture is built for — plus solid grounding and counting precision. It trails larger comparison models on knowledge-heavy STEM reasoning (MMMU) and complex instruction-following, consistent with its position as a small, specialization-friendly foundation model rather than a general-knowledge frontier system.

Specifications

Cohere Labs' launch blog does not separately specify the model's token context window.

Pricing

Open weights, self-hostable — no per-token API pricing. Available on Hugging Face under the Apache 2.0 license; public vLLM support is noted as "coming soon" in the launch announcement. Designed to be practical to fine-tune for specialized visual domains and, with the right inference stack and quantization, to run beyond server-only deployments — including on laptops and edge/mobile-class hardware.

Public Benchmark Scores

Scores are self-reported by Cohere Labs in the official North Micro Vision Instruct launch post, evaluated with VLMEvalKit. RefCOCO score is a P@1 average over the RefCOCO_val/testA/testB, RefCOCO+_val/testA/testB, and RefCOCOg_val/test splits.

Notable Aggregate Results

Cohere Labs' launch blog also reports capability-group averages across the same benchmark suite, computed by Cohere Labs (methodology not independently re-derivable from the individual per-benchmark scores above, so shown here as reported rather than recomputed):

• Overall average — 63.3%, ranking second among the six compact models compared: Qwen3.5-2B (Instruct Mode) leads at 68.2%, followed by North Micro Vision Instruct (63.3%), LFM2.5-VL-1.6B (61.6%), Ministral-3-3B-Instruct (60.5%), Phi-3.5-vision-instruct (59.8%), and gemma-4-E2B-it (53.0%)
• General VQA average — 54.6%
• Multilingual average — 68.2% — the strongest capability group in the comparison set
• Multi-image average — 52.7%
• Chart/Document/OCR average — 70.2%
• Grounding/Counting average — 72.9%
• Hallucination (HallusionBench) — 61.5%

North Micro Vision Instruct vs Alternatives

Despite being one of the smallest models in the comparison alongside Qwen3.5-2B, North Micro Vision Instruct posts the strongest DocVQA and RefCOCO-avg grounding scores, while Phi-3.5-vision-instruct keeps a narrow edge on ChartQA. All comparison scores are as reported by Cohere Labs in the same launch benchmark table; see each model's own page for scores from other sources.

FAQ

Is North Micro Vision Instruct open source? Yes — available on Hugging Face under the Apache 2.0 license, downloadable, fine-tunable, and deployable without restrictions.

How big is North Micro Vision Instruct? 2.4 billion total parameters, combining a custom 400M-parameter native-resolution vision encoder with an in-house 2B-parameter "North Micro LLM."

What is North Micro Vision Instruct best used for? Cohere Labs positions it as a compact foundation for specialized multimodal applications — document and chart understanding, multilingual visual QA, and visual grounding/counting — designed to be practical to fine-tune for specific visual domains and deployable on edge or mobile-class hardware.

What vision encoder does North Micro Vision Instruct use? A custom native-resolution vision encoder continued-pretrained from the SigLIP2 SO400M checkpoint (google/siglip2-so400m-patch16-384), using Continual Rotary Position Embedding (C-RoPE) to support variable, native-resolution inputs up to a full A4 page at 200 dpi.

Does North Micro Vision Instruct support reasoning/chain-of-thought? The launch blog does not describe it as a reasoning model with extended chain-of-thought output; it is positioned as a compact, direct-answering vision-language model.

Benchmark scores sourced from Cohere Labs' official North Micro Vision Instruct launch announcement (huggingface.co/blog/CohereLabs, Aug 12, 2026) and the CohereLabs/North-Micro-Vision-Instruct model card on Hugging Face. Comparison-model scores (Ministral-3-3B-Instruct, LFM2.5-VL-1.6B, Phi-3.5-vision-instruct, gemma-4-E2B-it, Qwen3.5-2B) are as reported in that same launch table.