Cohere Labs debuts North Micro Vision, 2.4B native-res VLM

Cohere Labs quietly shipped its smallest vision-language model to date this week, and the interesting part is not that it exists but where the tradeoffs land. North-Micro-Vision-Instruct, posted on Hugging Face under Apache 2.0, is a 2.4B-parameter open-weight VLM that pairs a custom-trained 400M-parameter vision encoder with a 2B-parameter in-house language model built on Cohere's Command A+ architecture.

The pitch is native resolution. Where most small VLMs downsample images to a fixed grid, North Micro accepts inputs up to 1654 by 2339 pixels, which Cohere describes as "an A4 page at 200 dpi", and preserves the aspect ratio. That framing is aimed squarely at document work, and the published numbers reflect it: 0.921 on DocVQA (ahead of Ministral-3-3B-Instruct at 0.896 and Phi-3.5-vision-instruct at 0.860) and 0.808 on ChartQA, where Phi-3.5-vision at 0.821 still edges it out.

The picture is more mixed elsewhere. On CountBench, North Micro scores 0.725, trailing LFM2.5-VL-1.6B at 0.910 and Ministral-3-3B at 0.737. On text-only MMLU it lands at 0.504, well behind Ministral-3-3B's 0.660. Cohere calls it "our smallest VLM to date" and pitches it as "a compact foundation for specialized multimodal applications", which reads as honest positioning for a model that leads on documents but not on counting or general knowledge.

Several practical details are missing. The post does not publish latency or throughput figures for the native-resolution mode, and there is no mention of a hosted API or pricing. Public vLLM support is marked as coming soon, not shipped. Community quantizations from Prince Canuma and Neywa are already available via MLX-VLM, and NVIDIA has published an AutoModel fine-tuning recipe alongside Axolotl support. Two experts in our Who's Who directory have already circulated the announcement.

For teams paying per-page to hosted document APIs, a permissively licensed 2.4B model that ingests a full A4 page in one shot is worth benchmarking on your own PDFs before renewing that contract. Just do not extend the document-benchmark wins into workloads Cohere's own tables show it losing.