North Micro Vision Instruct review, pricing and limits

In Cohere's own 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 SmolVLM2.2B), it scores 92.1% on DocVQA (VAL), half a point behind Qwen3.5-2B's 92.6%, and 73.2% on the RefCOCO grounding average. That grounding figure is more than double four of the seven peers: Ministral-3-3B at 31.7%, Gemma-4-E2B at 8.4%, Qwen3-VL-2B at 30.4% and SmolVLM2.2B at 1.8%. For readers shopping the same size class, the Ministral 3 3B and LFM2.5-VL-3B records show how neighbouring open models are specified. It is weakest on STEM and general knowledge: 32.9% on MMMU (DEV_VAL) is the lowest of all eight models in the table, and 50.4% on MMLU sits well behind Gemma-4-E2B's 69.2% and Ministral-3-3B's 66.0%.

The language backbone nominally supports a 128,000-token context window, but Cohere validates image-plus-text prompts only up to 8,000 tokens; longer multimodal contexts rely on extrapolation that has not been benchmarked. The vision encoder processes images at native resolution up to 1654 by 2339 pixels, the size of an A4 page scanned at 200 dpi, and keeps the aspect ratio. Input is interleaved text and multiple images, output is text only, and documented skills include visual question answering, captioning, grounding with boxes on a normalised 0-1000 scale, OCR, and chart and document reading across roughly a dozen languages. Tool calling and agentic workflows are not supported.

The weights are Apache 2.0 and published on Hugging Face. There is no metered Cohere API and no published hosted price for this checkpoint, so the only cost is compute. The model card documents Transformers (the card pins version 5.16.0), and community builds cover llama.cpp, Ollama, LM Studio, Jan and MLX for Apple Silicon. Cohere's launch post still lists public vLLM support as coming soon, so treat vLLM as unconfirmed. Fine-tuning recipes exist for NVIDIA NeMo AutoModel, plus community support in Axolotl. For a wider walk through local setups, see the guide to running AI locally and the Ollama vs LM Studio comparison.

Training ran in four stages: three resolution-scaling pretraining phases of 10 million, 13 million and 10 million examples, growing from 384 by 384 pixels to the full native cap, then a 50-million-example instruction-tuning stage weighted toward OCR (17.8%), charts and tables (17.8%), grounding (13.3%) and visual question answering (11.2%), and a final 500,000-example preference stage aimed at safety and formatting. Cohere has not published a HarmBench or jailbreak-resistance score, a training-data cutoff or a dedicated system card for this checkpoint, and it advises against system prompts because the model was not trained with one.

Who should look elsewhere: anyone who needs a general chat assistant, tool use, or strong STEM and math reasoning. Qwen3.5-2B beats it on MMBench (76.0% against 68.7%), MMStar, OCRBench and RefCOCO (78.5% against 73.2%), and it also leads on OCRBench v2, where this model's 36.7% trails five of the seven peers. Teams that outgrow this size can compare larger open models such as Gemma 4 12B. A hosted OCR service such as Mistral OCR 4 suits teams that do not want to run any model themselves. The model's own case is narrower: local invoice, receipt and form extraction, and a small base to fine-tune for one visual domain. To compare it with other open models, browse the model directory or ask Smart Match with your document type and hardware.