DeepSeek’s experimental vision model tops NVIDIA on one shared AI index

DeepSeek’s experimental V4 Flash Vision Exp scored 34.84 on the Artificial Analysis Intelligence Index, ahead of NVIDIA’s Nemotron 3.5 Super VL at 26.49, according to OrcaRouter. The comparison’s more useful point is that these are not equivalent products: one is an experimental release, the other a general-availability model with a NIM service.

Watch Desk analysis

What happened

OrcaRouter’s comparison says both scores are non-estimated results from the same index. It also reports that the DeepSeek configuration costs $0.44 per million input tokens and $1.32 per million output tokens, while NVIDIA has published no price for Super VL. The DeepSeek model is available as weights under MIT, but OrcaRouter describes it as an experimental multimodal release with a minimal reference implementation. It says NVIDIA’s model is generally available, has a technical report and is offered through NIM.

The comparison identifies different strengths, too. DeepSeek’s model scored higher on the index, while NVIDIA’s covers text, images and video and has published deployment guidance. OrcaRouter says the index result for DeepSeek is from a served configuration, not a configuration DeepSeek sells. These details come from OrcaRouter’s comparison, which draws on Artificial Analysis and the models’ published materials.

Why it matters

A higher score is useful evidence, not a universal buying instruction. The index offers a shared comparison, but a team choosing a model also needs to weigh whether it can deploy the model, what it costs, which inputs it handles and what support comes with it. Here, experimental weights and a supported service are quite different propositions, even when the benchmark puts them on the same chart.

OrcaRouter also flags that vendor benchmark tables use different tests and setups, so their figures should not be treated as directly comparable. One common index is a better starting point than a pile of self-selected scores, though it cannot answer every deployment question.

Our read

The headline score belongs to DeepSeek. The practical choice is less tidy: an experimental model currently leads this index, while NVIDIA’s offering looks more packaged for deployment. That is an interesting result, not a verdict on which model will work better for your workload. Test the task you actually need done, and check whether the tested configuration is the one you can use. The benchmark is a map, not the territory, and occasionally the territory has an invoice.

What to watch

• Whether DeepSeek turns the experimental model into a supported, deployable product.
• Whether Artificial Analysis updates the index or its model configurations.
• Whether NVIDIA publishes pricing for Super VL.
• How both models perform on the same real-world workloads.

Discussion spark: Would you choose the higher-scoring experimental model, or the more deployment-ready option when the benchmark gap is clear but the products are not like for like?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.