Theme
AI Resources
North Micro Vision Instruct
North Micro Vision Instruct is Cohere Labs' 2.4B-parameter vision-language model for reading and answering questions about documents, charts, tables, forms, screenshots, and natural images.
It keeps image aspect ratios and fine detail through native-resolution processing, accepts multiple images and several languages, and provides downloadable weights for prototyping or task-specific fine-tuning. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
A small model for detailed visual inputs
The model combines a custom 400M vision encoder with a 2B language model. It accepts interleaved text and images and returns text, including normalized bounding-box coordinates for visual grounding.
Why it stands out
Documents keep their shape and fine detail
Instead of forcing every input into one small square, the vision encoder preserves varied aspect ratios and can process a single A4 document page at up to 200 dpi. That makes small text, layout, tables, and chart details central to the design.
Availability
Download, adapt, or fine-tune the model
The main path uses Hugging Face Transformers. Cohere also links an NVIDIA AutoModel recipe, Axolotl support, and community MLX weights for builders adapting the model to a visual domain or deployment target.
Why it matters
What makes it useful
A compact model that keeps document detail can be adapted for jobs such as extracting fields from forms, reading charts, locating items in an image, or answering questions across several pages or pictures. Its size makes experimentation and focused fine-tuning more approachable than starting with a much larger general assistant.
What to know
Where it fits
The model is aimed at research, prototyping, and specialized multimodal applications. It fits best when visual understanding or structured extraction is the main job; Cohere says it is not a reasoning model, has limited math and code ability, and does not support tool calling or agentic workflows.
Notable points
What stands out
Cohere reports strong results in parts of its document, chart, OCR, and grounding evaluation, but the detailed table is mixed rather than a clean win across every task. Treat those publisher-run benchmarks as a starting point and test the exact document types, languages, image sizes, and output format the workflow needs.
Before using
What to review
The model card currently requires Transformers 5.16.0; until that release is on PyPI, its setup instructions install Transformers from source.
Public vLLM support is marked as coming soon, so use the documented Transformers route unless the current model card says otherwise.
Multimodal training used an 8K-token context. The language backbone supports more, but longer multimodal prompts were not benchmarked.
Native-resolution inputs can use more memory and take longer as image dimensions grow. Test representative scans, screenshots, charts, and multi-image prompts on the intended hardware.
Check extracted text, tables, labels, coordinates, and structured fields against the source before they enter search, records, or automation.
System prompts are not recommended because the model was not trained with them, even though its chat template accepts the system role.
The model card names Apache 2.0. Review the current terms at the main model page for the intended use.
Reader fit
Who may find it relevant
Builders prototyping OCR, document extraction, chart reading, image questions, captioning, or visual grounding with downloadable weights.
Teams fine-tuning a compact visual model for a particular document type, language mix, or specialist image domain.
People comparing whether native-resolution inputs preserve useful detail in forms, tables, scientific figures, screenshots, and scanned pages.
Less relevant for readers who need a ready consumer app, deep reasoning, coding, tool use, or an established vLLM serving path today.
Editorial note
Why LifeHubber lists it
North Micro Vision Instruct makes native-resolution document and chart work available in a 2.4B model that builders can download and adapt. LifeHubber lists it to help readers judge whether a smaller specialized model can preserve the visual detail their extraction or question-answering workflow depends on.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
What to explore next
Connect the model to the document workflow around it.
The model handles visual inputs, but the wider decision includes how it compares with other model jobs, where it runs, and how extracted answers stay connected to the original material.
More in AI Models
Keep browsing this category
Explore more AI model resources.
Gemma 4
google/gemma-4
A Google DeepMind Gemma 4 model family collection with public checkpoints including Gemma 4 12B, a dense multimodal model Google describes around local agentic workflows, native audio input, and encoder-free vision/audio handling.
DeepSeek-OCR-2
deepseek-ai/DeepSeek-OCR-2
A newer DeepSeek OCR model release for image/PDF OCR, document-to-Markdown workflows, dynamic resolution, vLLM/Transformers inference, and visual causal flow research.
MiniMax-M2.7
MiniMaxAI/MiniMax-M2.7
A large MiniMax model focused on agentic work, software engineering, tool use, and complex productivity workflows.
