Released August 12, 2026, North Micro Vision Instruct pairs a 128K-token language backbone with an 8K-token validated multimodal range, aimed at local document and OCR pipelines rather than general chat. It's the vision debut of Cohere's new open-weight North family, built for developers who want to self-host instead of calling a metered API.
North Micro Vision Instruct is a compact open-weight vision-language model from Cohere Labs, scoring 73.2% on RefCOCO visual grounding, well clear of most same-size open peers. It reads full-page scans and photos at native resolution without downscaling, and ships with fully open, unrestricted licensing.
Provider: Cohere · Family: North
Context window: 128,000 tokens
Input modalities: text, image · Output: text
About North Micro Vision Instruct
North Micro Vision Instruct is a 2.4-billion-parameter vision-language model released by Cohere Labs, the open-science arm of Cohere, on August 12, 2026. It pairs a custom 400-million-parameter native-resolution vision encoder, continued-trained from Google's SigLIP 2 SO400M, with a 2-billion-parameter North Micro language model built on Cohere's Command A+ architecture. It is the second entry in Cohere's new North family of workstation-scale open models, following North Mini Code, a 30B-parameter agentic coding model released June 9, 2026, and the first to add vision. Cohere positions it explicitly as a compact foundation for prototyping and fine-tuning, not a general chat-assistant replacement. On Cohere's own published comparison against seven similarly sized vision-language models (Ministral-3.3B, LFM2.5-VL-1.6B, Phi-3.5-vision-instruct, Gemma-4-E2B, Qwen3-VL-2B, Qwen3.5-2B, and SmolVLM-2.2B), North Micro Vision Instruct scores 92.1% on DocVQA (VAL), within half a point of the class-leading Qwen3.5-2B's 92.6%, and 73.2% on RefCOCO average grounding, more than double five of those seven peers (Ministral-3.3B 31.7%, Gemma-4-E2B 8.4%, Qwen3-VL-2B 30.4%, SmolVLM-2.2B 1.8%). It trails badly on STEM and general knowledge: 32.9% on MMMU_DEV_VAL, the lowest of all eight models compared, and 50.4% on MMLU, well behind Gemma-4-E2B's 69.2% and Ministral-3.3B's 66.0%. The North Micro language backbone nominally supports a 128,000-token context window and a 262,144-token vocabulary, but Cohere validates multimodal image-plus-text prompts only up to 8,000 tokens; longer multimodal contexts rely on unbenchmarked extrapolation. Architecturally, the language model interleaves three sliding-window attention layers using rotary positional embeddings with one global attention layer, and the vision encoder combines 2D RoPE with learned 1D positional embeddings to preserve spatial structure. Input is interleaved text and multiple images; output is text only. The vision encoder processes images at native resolution up to 1654 by 2339 pixels, equivalent to a full A4 page scanned at 200 dpi, without downscaling, and preserves aspect ratio. Documented capabilities span visual question answering, image captioning, visual grounding with bounding boxes returned on a normalized 0-1000 coordinate scale, OCR, and chart and document understanding, across 12 or more languages including English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese, and Arabic. Tool calling and agentic workflows are not supported. North Micro Vision Instruct ships under the Apache 2.0 license with weights published on Hugging Face; there is no metered vendor API and no published Cohere-hosted pricing for this checkpoint. The only cost is self-hosting compute, and its 2.4B-parameter size keeps that modest next to larger vision-language models built for the same document and OCR tasks. Cohere evaluated the model internally with vLLM, with public vLLM support coming soon at release. Supported today: Hugging Face Transformers from source (version 5.16.0 or newer, with optional Flash Attention 2), llama.cpp, Ollama, LM Studio, and Jan via community GGUF builds, and MLX for Apple Silicon via community-contributed weights. Fine-tuning recipes exist for NVIDIA's NeMo AutoModel framework and for Axolotl, community-supported. Training followed a four-stage pipeline: resolution-scaling pretraining stages of 10 million, then 13 million, then 10 million examples, growing from 384x384 to full 1654x2339 native resolution, a 50-million-example instruction-tuning stage weighted toward OCR (17.8%), charts and tables (17.8%), grounding (13.3%), and VQA (11.2%), and a final 500,000-example stage using a simplified Mixed Preference Optimization method focused on safety and response formatting. Cohere has not published a HarmBench or jailbreak-resistance score for this checkpoint, and system prompts are explicitly discouraged since the model was not trained with one, even though its chat template technically accepts a system role. Best suited to developers building local or on-device document, invoice, receipt, and OCR extraction pipelines, teams fine-tuning a compact VLM for a narrow visual domain rather than open-ended chat, and researchers who need strong visual grounding without a large model footprint. Teams needing a general-purpose chat assistant, agentic tool use, or strong STEM and math reasoning should look at Qwen3.5-2B, which beats North Micro Vision Instruct on MMBench, RefCOCO, and MMLU-Pro simultaneously, or a larger multimodal model entirely. Cohere has not published an exact training-data cutoff date or a dedicated system card for this checkpoint at time of writing; the Hugging Face model card is the primary documentation. As an open-weights, self-hosted model, no user data is sent to Cohere at inference time unless a third-party hosted-inference provider is used. North Micro Vision Instruct is the first vision entry in Cohere's month-old North family, which launched with North Mini Code in June 2026. Cohere frames North as an extension of its enterprise North agent-platform brand down to the workstation, aimed at sovereign, locally deployable open-source models.
Pricing
No metered Cohere API for this checkpoint. The weights are free under an open license; the only real cost is self-hosting compute, which stays modest thanks to the model's compact size, and drops further with community quantized builds.
Key Features
- Native-Resolution Vision Encoder: Reads scanned pages and photos at their original size instead of a fixed square crop, keeping small text and fine chart detail legible.
- Strong Document Understanding: Scores among the top of its size class on document visual question answering, per Cohere's own published benchmark comparison.
- Reliable Visual Grounding: Locates and boxes objects in an image far more reliably than most similarly sized open models, per Cohere's RefCOCO results.
- Multilingual, Multi-Image Input: Understands interleaved text and multiple images together across a dozen languages, from English and Spanish to Japanese, Korean, and Arabic.
- Open and Locally Runnable: Ships as open weights with no commercial-use restriction and runs through popular community tooling for offline, local deployment.
Pros
- Matches the strongest same-size vision-language models on document understanding, per Cohere's own published benchmark comparison.
- Clears most similarly sized open peers by a wide margin on visual grounding, a benchmark most small VLMs struggle with.
- Fully open licensing with no commercial-use restriction, and broad community support for running it locally on modest hardware.
- Reads full-resolution scans and photos without downscaling, preserving fine text and chart detail that smaller, fixed-resolution encoders lose.
Cons
- MMMU_DEV_VAL of 32.9% is the lowest score among every model in Cohere's own published comparison table, so STEM visual reasoning is a clear weak point.
- MMLU of 50.4% trails Gemma-4-E2B and Ministral-3.3B by a wide margin on general text knowledge, since this checkpoint wasn't trained as a broad reasoning model.
- Multimodal inputs are only validated to a fraction of the language backbone's full context window, so very long documents or image sequences fall outside tested behavior.
- The chat template technically accepts a system-role message, but the model wasn't trained to rely on one, so results can be inconsistent if you use it. Tool calling and multi-step agentic workflows aren't supported either.
Benchmarks
- mmlu: 50.4
- mmmb: 72.8
- blink: 52.7
- ifeval: 74.9
- mmstar: 51.8
- mmlu pro: 30.7
- multi if: 37.3
- ocrbench: 79.2
- ai2d test: 77.5
- charxiv dq: 60
- countbench: 72.5
- docvqa val: 92.1
- infovqa val: 65.2
- realworldqa: 62.2
- refcoco avg: 73.2
- chartqa test: 80.8
- mmmu dev val: 32.9
- hallusionbench: 61.5
- ocrbench v2 en: 36.7
- mtl mmbench dev: 63.6
- mmbench dev en v11: 68.7
- gqa testdev balanced: 57.4
Frequently Asked Questions
How much does North Micro Vision Instruct cost to run?
North Micro Vision Instruct is free and open-weight, so there's no per-token API fee from Cohere. The only cost is self-hosting compute: the bf16 checkpoint needs roughly 4.8GB of VRAM, and community quantized builds bring that down to an estimated 1.5-2.5GB. This differs from Cohere's Command models, which are metered through Cohere's own commercial API.
How does North Micro Vision Instruct compare on benchmarks to Qwen3.5-2B?
Qwen3.5-2B, the strongest peer in Cohere's own comparison table, leads on general visual question answering (MMBench 76.0% vs North Micro Vision's 68.7%) and multilingual tasks (MMMB 74.5% vs 72.8%). North Micro Vision Instruct stays close on document and grounding benchmarks and adds native-resolution image input that Qwen3.5-2B doesn't offer at the same parameter count.
Is North Micro Vision Instruct open source or proprietary?
It's fully open-weight under the Apache 2.0 license, published by Cohere Labs on Hugging Face. There are no commercial-use restrictions, and the weights can be fine-tuned, quantized, and redistributed under the same terms.
Does North Micro Vision Instruct train on user data?
No. It's a self-hosted, open-weight model, so nothing run through it is sent to Cohere unless a third-party hosted-inference provider is chosen. Cohere hasn't published a dedicated data-retention policy for this checkpoint, since there's no vendor-side inference to retain in the first place.
Who is North Micro Vision Instruct best for and who should avoid it?
It's a good fit for building local document, invoice, or OCR extraction tools, and for fine-tuning toward one specific visual task rather than open-ended chat. Anyone wanting a general assistant, agentic tool use, or strong math and STEM reasoning should pick a larger multimodal model instead.