NVLM-D 72B on live video

What NVLM-D 72B is good at

NVLM-D 72B uses a dynamic high-resolution tiling scheme with explicit tile tags, so the model can track which part of a large image each tile came from rather than losing spatial context when an image is split up for encoding. That underpins strong performance on OCR-heavy content and mathematical reasoning over diagrams and equations.

A notable property of NVLM-D 72B is that its text-only performance actually improves over its Qwen2-72B-Instruct backbone after multimodal training, rather than the usual tradeoff where adding vision capability costs some language quality. That makes it a reasonable single model for mixed text-and-vision workloads.

• OCR and math reasoning over diagrams, equations, and dense documents
• High-resolution image understanding via dynamic tiling with tile tags
• Combined text and vision workloads without a language-quality tradeoff

NVLM-D 72B and live video workflows

NVLM-D 72B is not currently in Overshoot's live model catalog; GET /v1beta/models returns the models that are live at any given time. The workflow it would plug into is standard across the API: publish a camera or screen share over WebRTC to open a Stream, then send a chat-completions request whose image_url or video_url is an ovs:// reference, anchored to the latest frame, an exact timestamp, or a recent segment.

NVLM-D 72B's tiling approach is well suited to reading dense on-screen detail from a live feed, such as a whiteboard or a document under a camera, so it is the kind of model that maps naturally onto that streaming pattern if it enters the catalog.

NVLM-D 72B and licensing versus similar open models

NVLM-D 72B sits in the same size class as InternVL3 78B and Qwen2.5-VL 72B, but its CC-BY-NC-4.0 license restricts it to research and non-commercial use, which is the key factor to weigh against those alternatives when picking a model for a production deployment. Where non-commercial use is acceptable, NVLM-D 72B’s tiling and OCR strengths make it a strong research baseline.

Frequently asked questions

Can NVLM-D 72B analyze live video?

NVLM-D 72B accepts images and video frames, so it can analyze video when self-hosted for research use. It is not currently in Overshoot's live model catalog, so it cannot be referenced against an Overshoot WebRTC stream today. The catalog changes over time; GET /v1beta/models lists what is live.

Is NVLM-D 72B free to use commercially?

No. NVLM-D 72B is released under CC-BY-NC-4.0, a non-commercial research license, so it is best suited to research, evaluation, and internal prototyping rather than a commercial production deployment.

Is NVLM-D 72B available on Overshoot?

Not currently. NVLM-D 72B is not in Overshoot's live model catalog. The catalog changes over time, so check GET /v1beta/models for the up-to-date list of hosted and passthrough models before building against a specific model.

How does NVLM-D 72B compare with InternVL3 78B?

Both are 70B-class open vision-language models with strong OCR and reasoning ability. InternVL3 78B carries a more commercially permissive weight release, while NVLM-D 72B is restricted to non-commercial use under CC-BY-NC-4.0, which is usually the deciding factor between them.

Related models

• InternVL3 78B
• Qwen2.5-VL 72B
• Molmo 72B
• Llama 4 Scout