Cohere releases North Micro Vision, a 2.4B open-weight vision-language model

Cohere's research arm, Cohere Labs, published North Micro Vision on August 12, 2026, a compact open-weight vision-language model built to preserve fine visual detail — tables, charts, screenshots, and forms — at native resolution rather than downsampling images the way many small VLMs do.

What's new

Cohere Labs describes the release plainly: "A 2.4B-parameter open-weight vision-language model (VLM) with native-resolution image support." Under the hood, the model pairs a "400M-parameter native-resolution vision encoder" with a "2B-parameter language model," and ships under the "Apache 2.0 license" — a fully permissive license that allows commercial use, modification, and redistribution without royalty.

The model is built around "dense, native-resolution visual inputs" that preserve the "aspect ratio and fine detail of documents, tables, charts, screenshots, and forms" — a deliberate design choice aimed at document-heavy use cases where downsampling an image tends to destroy the small text or fine lines that make it useful in the first place. Cohere Labs also highlights "multilingual visual understanding" spanning multiple languages and visual domains, along with strengths in document understanding and visual grounding. In partnership with NVIDIA, Cohere is shipping an AutoModel recipe alongside the release, so developers can fine-tune and deploy the model directly on NVIDIA GPUs.

Context

North Micro Vision extends a lineup Cohere Labs has been building out through 2026, following releases like North Mini Code, its first developer-focused coding model, earlier in the year. Where Cohere's enterprise business has increasingly focused on larger, workflow-oriented products — including North Automations, a workflow-orchestration layer launched for its enterprise platform in late July — North Micro Vision is explicitly positioned at the small end of the spectrum: a "compact foundation for adaptation to domain-specific data" rather than a general-purpose frontier multimodal model competing directly with GPT or Gemini's vision capabilities.

That positioning matters in a market where most attention goes to ever-larger frontier vision-language models. A 2.4B-parameter model that keeps native image resolution is aimed squarely at teams that need to fine-tune a VLM for a narrow task — parsing a specific document format, reading a particular kind of chart or form — without the infrastructure cost of serving or customizing a multi-billion-parameter frontier model.

Why it matters

Open-weight, Apache-licensed small VLMs lower the barrier for teams building document-processing and visual-QA pipelines that need to run outside a hyperscaler API — on-premises, at the edge, or fine-tuned tightly enough to a narrow domain that a general frontier model would be both overkill and less accurate. The explicit design tradeoff — sacrificing broad frontier-level capability for native-resolution fidelity at a fraction of the parameter count — targets a specific, underserved niche: OCR-adjacent and document-understanding workloads where losing fine visual detail to downsampling is a correctness problem, not just a quality one. Shipping a ready-made NVIDIA AutoModel recipe alongside the weights also lowers the deployment friction that often keeps smaller open models from seeing real production use.