Cohere Labs ships 2.4B open-weight vision-language model

Cohere Labs has posted a small vision model to Hugging Face that is quietly interesting for teams building multimodal features on constrained hardware. The North Micro Vision Instruct model card describes a 2.4B-parameter open-weight vision-language model released under Apache 2.0, pairing a 2B-parameter language backbone with a 400M-parameter vision encoder that Cohere says was custom-trained starting from SigLIP 2 SO400M.

The framing is unusually modest. Cohere positions the model as a compact foundation for prototyping, task-specific fine-tuning, and specialized multimodal applications rather than a general-purpose chat assistant. Native-resolution image processing preserves aspect ratios and fine visual detail, aimed at VQA, captioning, grounding, OCR, and chart and document understanding. The language stack advertises a 128K-token context window and covers English, German, French, Spanish, Italian, Portuguese, Hindi, Japanese, Korean, Chinese, Arabic, and more.

Read the limitations closely. Multimodal training used an 8K-token context, and Cohere writes that longer contexts have not been validated. The card states plainly that the model is not a reasoning model and has limited math and code-generation capabilities, that tool calling and agentic workflows are not supported, and that system prompts are not recommended because the model was not trained with them. Cohere's own comparison table reflects that positioning: North Micro scores 0.329 on MMMU, well below Ministral-3-3B-Instruct's 0.508 and Qwen3.5-2B's 0.474. Where it does hold its own is document and chart work, with 0.921 on DocVQA, 0.808 on ChartQA, and 0.732 on RefCOCO grounding.

The card does not publish training data provenance, inference latency figures on common accelerators, or a safety evaluation, so anyone shipping this into production still has due diligence in front of them. The Apache 2.0 license is the actual story here: three tracked researchers in our Who's Who directory shared the link, and the specific move is that a frontier lab has put a permissive, native-resolution multimodal base into the open ecosystem for people to fine-tune. If you are building a narrow document, OCR, or chart pipeline and have been priced out of hosted VLMs, this is a plausible starting point.