Nemotron 3.5 Super VL vs DeepSeek V4 Flash Vision Exp: Eight Index Points, and Only One of Them Is a Product

On the Artificial Analysis Intelligence Index, DeepSeek V4 Flash Vision Exp beats NVIDIA Nemotron 3.5 Super VL by 8.35 points — 34.84 to 26.49, both non-estimated — and it does so at a published $0.44 per million input tokens and $1.32 output while NVIDIA has published no price at all. That is the summary a spec-sheet reader walks away with, and it is the wrong one to act on. The two models are not competing for the same slot. NVIDIA Nemotron 3.5 Super VL is a general-availability model with a technical report, a NIM microservice, enterprise support and a post-training ecosystem that includes Scale AI, Salesforce and CrowdStrike. DeepSeek V4 Flash Vision Exp is, by its own filename, an experiment: the first multimodal member of the V4 family, shipped as weights plus a minimal PyTorch reference implementation, and scored on that panel in a configuration DeepSeek does not sell. Eight points is a real gap. It is also a gap between a service and a research artifact, and the useful question is not who wins but which one can sit inside your compliance boundary in six months.

What each one actually is

NVIDIA Nemotron 3.5 Super VL is a 120B-parameter hybrid — Mamba-2 layers, mixture-of-experts with 512 routed experts, and attention — with a C-RADIO vision encoder, 12B active parameters per token, up to 1M tokens of context and native Multi-Token Prediction for speculative decoding without a separate drafter. Its inputs are text, images and video; output is text. NVIDIA's own evaluation covers OCR, chart and document reasoning, video understanding and long-context work, and the licence is OpenMDW-1.1, which permits commercial use. It carries a technical report and a published deployment recipe for 4×H100 or 2×GB200.

DeepSeek V4 Flash Vision Exp builds on the DeepSeek-V4-Flash architecture, adding visual modules and continued training. The repository carries 48 safetensors shards under MIT, a tokenizer, an encoding module that converts OpenAI-style message blocks into prompts, and a minimal PyTorch implementation covering the vision encoder, aligner, DFlash attention, MoE, Hyper-Connections and the DSpark forward path. The config shows a 43-layer MoE with 256 routed experts, 6 experts per token, 1,048,576 maximum positions and a three-layer multi-token-prediction head. The card is explicit that this is DeepSeek's first experimental multimodal model in the V4 line, and that it improves on DeepSeek-V4-Flash-0731 on multimodal agent tasks while holding comparable performance on text-only ones.

• Parameters and shape — NVIDIA Nemotron 3.5 Super VL: 120B total / 12B active hybrid MoE with C-RADIO vision. DeepSeek V4 Flash Vision Exp: 284B total / 13B active MoE, 256 routed experts, 6 per token.

• Intelligence Index — 26.49 for Super VL, 34.84 for the served DeepSeek configuration. Both non-estimated, both per Artificial Analysis.

• Output tokens per index task — 49,174 against 69,174. Super VL finishes the panel's tasks with 29% fewer tokens.

• Price — no published rate anywhere for Super VL; $0.44 in / $1.32 out per million for the served DeepSeek configuration, per the same panel.

• Licence and commitment — OpenMDW-1.1 with commercial use, GA and a NIM service for Super VL; MIT weights and a reference implementation, labelled experimental, for DeepSeek V4 Flash Vision Exp.

• Modalities — text, image and video in for Super VL per NVIDIA's card; image in, text out for the DeepSeek model, whose own benchmark table has no video rows at all.

The eight-point gap holds up; the price comparison does not

Start with the part that is genuinely settled. DeepSeek's model scores higher and costs less per token, and the margin is not noise — 8.35 index points is roughly the distance between NVIDIA Nemotron 3.5 Super VL and NVIDIA's own 550B NVIDIA Nemotron 3 Ultra 550B A55B, which sits at 22.93. So the honest framing is that a Chinese lab's experimental 284B multimodal release currently out-scores an American lab's flagship open vision model, at roughly a third the size on the active-parameter axis and at a published rate. If your workload is screen and chart understanding delivered over an API, that is the whole answer and you can stop reading.

The comparison breaks down the moment you want to run it yourself, which is the reason both of these models exist. The 34.84 figure is attached to the served Max configuration, not to a checkpoint you can deploy. The Hub repository is a genuine weight release under MIT, but DeepSeek's own card presents it as an experimental artifact with a minimal inference path — the kind of release a lab publishes to let researchers reproduce a paper, not the kind it supports with an enterprise contract. NVIDIA is doing the opposite: publishing weights at the same time as the product, with the caller expected to host. On the benchmarks the two labs choose to publish, the difference shows. DeepSeek emphasises multimodal agent work — ApexBench 36.5 against its own 26.2 for the text-only model, Chartography 64.3, ZeroBench at 35.0 pass@5, Agents' Last Exam 27.3. NVIDIA emphasises document and video surfaces — OCRBench v2 at 68.12 in English, MMLongBench-Doc 62.82, CharXiv reasoning 80.82, Video-MME v2 32.85. Neither set is comparable to the other, and both are vendor-reported.

Where the vendor numbers disagree with each other

There is a detail in NVIDIA's own table that is worth a paragraph, because it is exactly the kind of thing that gets laundered in aggregation. NVIDIA's comparison column includes DeepSeek-V4-Flash0731 and reports Terminal-Bench 2.1 at 76.80 for it. DeepSeek's own card reports 82.7 for the same model on the same benchmark, and 83.9 for DeepSeek V4 Flash Vision Exp. Three numbers, one benchmark, two harnesses, and no reconciliation on either side. This is not a scandal — different scaffolding and reasoning-effort settings move agentic benchmarks by several points, and NVIDIA publishes its NeMo Gym recipes precisely so the settings can be inspected. But it means that when you see an eight-point index gap, the index itself is the reliable part and the component benchmarks are not. The index is a single harness applied to both models by a third party. The benchmark tables are four labs grading their own homework on their own scaffolding.

On the same panel, both models carry a negative omniscience index. Super VL sits at minus 4.22 with 23.3% accuracy and a 35.8% hallucination rate; that row measures whether a model declines to answer when it does not know, and neither of these two is good at it. GDPval-AA — finished deliverables rather than answers — gives Super VL 1,162.41 and the AA-Briefcase rubric an Elo of 998.76. For the DeepSeek model the same rows are not populated in the served configuration. If your use case is knowledge retrieval, neither number is encouraging. If it is document production, Super VL has the only measured row in the pair.

Running either one

Deployment is where the paths separate for good. NVIDIA Nemotron 3.5 Super VL serves BF16 on 4×H100 or 2×GB200 with pinned vLLM and SGLang containers, adds an NVFP4 build for tighter footprints, and documents a video pipeline sampling 256 frames at 2 FPS with a 0.5 pruning rate. DeepSeek V4 Flash Vision Exp publishes a vLLM recipe for a single 4×GB300 node with FP8 KV cache, block size 256 and DSpark speculative decoding drawn from the same checkpoint, plus an SGLang path — lighter on the selling side, heavier on assumption that you know what you are doing. Both need datacentre-class hardware; neither is a laptop model. Notably, the one thing they share is that a vendor cut to list price shows up at cost on OrcaRouter the same day it lands, because we pass provider pricing through at 0% markup across 200-plus models on a single key, and automatic failover lets you put an unproven checkpoint on a slice of traffic while a proven one holds the rest. That is the practical answer when the eight-point gap and the missing price make the decision genuinely close: route both, measure on your own inputs, and let the traffic settle it — with the caveat that only one of the two can be routed today. DeepSeek V4 Flash Vision Exp is already on our catalogue, at the rate its provider publishes rather than a marked-up one; NVIDIA Nemotron 3.5 Super VL is not there yet, because nobody is reliably serving a three-day-old checkpoint. That asymmetry is worth saying out loud instead of implying we carry both.

The choice in one line each

Pick NVIDIA Nemotron 3.5 Super VL when the work is documents, charts, scanned pages and video, when the licence has to permit commercial use, and when you need a vendor who will still be answering support tickets about the checkpoint in a year — accepting that you are giving up 8.35 index points and paying for GPUs instead of tokens. Pick DeepSeek V4 Flash Vision Exp when the work is chart and screen understanding over an API, when $0.44 and $1.32 per million tokens is the budget line, and when the beta-grade framing of an experimental release is acceptable — accepting that the served configuration is not the checkpoint and the checkpoint is not supported. Teams that need both behaviours are usually running two different workloads, and two different workloads is what one endpoint with failover is for.