Cohere Releases North Micro Vision for Native-Resolution Documents
The compact open-weight model keeps document layouts and fine detail at native resolution, but its capability evidence and deployment readiness remain first-party and incomplete.
Cohere released North Micro Vision on August 12, 2026 as a 2.4-billion-parameter open-weight vision-language model under the Apache 2.0 licence. The model is aimed at developers adapting multimodal systems for documents, charts, screenshots and forms, where resizing every image to a small square can discard layout and fine detail.
The material change is not simply another compact VLM. North Micro Vision accepts native-resolution inputs and preserves their aspect ratio, giving document-AI and OCR teams a foundation designed around dense pages rather than treating them as ordinary photographs. Cohere has made the weights available, although the supporting deployment stack is not yet complete.
Native resolution is the core design choice
The model combines a custom-trained 400-million-parameter vision encoder with Cohere's in-house 2-billion-parameter North Micro language model. A projector connects their representations so the language component can use visual features extracted at several levels.
Cohere trained the visual path through a resolution curriculum. It began with fixed 384-by-384 images, then increased the working resolution and ultimately trained a checkpoint that can process one A4 page at up to 200 dots per inch while retaining its aspect ratio. The later instruction-tuning mix emphasised OCR, charts, tables, visual grounding and counting, while retaining text-only examples.
That focus is practical for invoices, archives, reports and interface screenshots because characters, cell boundaries and page structure can be load-bearing information. It also narrows the claim: this is a document-oriented compact model, not evidence that a smaller VLM is uniformly stronger across visual reasoning tasks.
Cohere's results show a specialist, not a universal winner
The first-party benchmark table supports that narrower reading. North Micro Vision scored 0.921 on the listed DocVQA validation set and 0.732 on the averaged RefCOCO grounding evaluation. Those results place it near the stronger compact comparison models in the release's document and grounding tasks.
The same table shows weaker areas. Its reported MMMU development-and-validation score was 0.329, below every comparison score shown, while HallusionBench was 0.615 and did not lead the group. Even within OCR and document tests, different peers finished ahead on individual benchmarks. The useful signal is therefore the model's capability shape, not a claim of overall benchmark leadership.
All of these measurements come from Cohere. The team used an internal vLLM implementation for North Micro Vision and the recommended settings for comparison models where available. No independent reproduction was included, so the scores should guide further testing rather than substitute for evaluation on a developer's own documents and languages.
What developers can use now
The model weights are public under Apache 2.0, which gives teams a concrete base for adaptation rather than an API-only preview. Cohere says the 2.4-billion-parameter scale is intended to support specialised fine-tuning and experimentation beyond server-only environments.
Public vLLM support was still marked as coming soon in the release. The announcement also does not provide latency, memory-use, quantisation-quality or serving-cost results for production hardware. Those omissions matter because a compact parameter count does not by itself establish that a document pipeline will meet a particular device, throughput or budget target.
What remains uncertain
North Micro Vision is a shipped open-weight release with a detailed training and evaluation description, but the evidence is still controlled by its maker. Independent benchmark runs, public vLLM support and workload-specific measurements are the next checks that would turn an interesting architecture into a clearer deployment decision.
The training description also refers to a large in-house multilingual document corpus without publishing enough detail here to audit its composition. Teams evaluating language coverage, historical documents or domain-specific forms will need to test those cases directly rather than infer them from the aggregate tables.
Status
Confirmed. Cohere has released the weights and documented the architecture, training process and internal evaluations. Internal confidence is medium because performance and deployment claims currently rely on the first-party release.
Sources
Update note: Last reviewed 2026-08-17. We will revise this article when public runtime support, independent evaluations or deployment measurements become available.
Sources
- Cohere Labs — Meet North Micro Vision — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.