Short version. NVIDIA says Nemotron 3.5 Super VL, a 120B model with 12B active parameters, runs on one DGX Spark. The model card agrees, but only for the NVFP4 checkpoint, whose files add up to about 75GB. That checkpoint's own benchmark table shows it trailing the BF16 build on a few tests by 3 to 11 points, and the Spark row of the deployment table lists text and images where the bigger setups list video too.
A 120B model on a desk. That's the pitch, and it's close to true. We pulled the two model cards, the file list and NVIDIA's launch post, and the Spark setup holds up with a few asterisks you'd want to know about before you order hardware around it.
What's actually in the repos
NVIDIA's launch post is dated 15 October, but the Super VL repos on Hugging Face were created on 12 October. Two builds matter. The BF16 build is the full-precision reference, and its card says to run it on 4 H100s or 2 GB200s. The NVFP4 build is the quantized one, and its card lists 1 DGX Spark, 1 H200, 2 H100s or 1 GB200. Both cards describe the same model: a Mamba-2 and Transformer hybrid mixture of experts, 120B total parameters, 12B active, context up to 1M tokens, text out and text, image or video in. The license on both is OpenMDW-1.1, which is why the Hugging Face tag just says "other".
We added up the file sizes of the NVFP4 repo through the Hugging Face API and got 75GB. That's the repo, not a measured memory footprint, so treat it as a floor for what has to sit in memory. A DGX Spark has 128GB shared between CPU and GPU, and the Spark recipe sets --gpu-memory-utilization 0.75. That leaves room on paper, though we haven't run it, so we can't tell you what the KV cache does at 1M tokens.
The Spark row has fine print
The deployment table on the NVFP4 card has one row per setup. The Spark row says vLLM, text and image, 1M-token context, speculative decoding with 5 draft tokens. The 2 H100, 1 H200 and 1 GB200 rows say text, image and video. So video is listed everywhere except the box people will buy for a desk.
Then it gets muddled. The quick-start recipe for the Spark, a few screens higher on the same card, sets --video-pruning-rate 0.5 and a video frame rate, which suggests video is meant to work. It also names a separate drafter checkpoint and asks for 7 speculative tokens, where the table says 5. It serves one sequence at a time, too: --max-num-seqs 1. Our read is that one of the two is stale, and I'd bet on the table. I might be wrong. Until NVIDIA fixes the card, test video on your own unit before you promise it to anyone.
One concurrent request is fine for a personal coding agent. It isn't a shared endpoint for a team.
What quantizing cost, in NVIDIA's own table
The benchmark table on the NVFP4 card isn't Super VL against rivals. It's BF16 against NVFP4, and NVIDIA says it held every other setting fixed. Most rows land within a point of each other. GPQA-Diamond goes up, from 89.52 to 89.90, and Terminal-Bench 2.1 goes from 60.67 to 62.78, which looks like noise, not a gain from fewer bits. SWE-bench Pro slips from 51.78 to 51.30.
Others don't hold. MMLU-ProX falls from 76.91 to 73.76. In the rows tagged [offline], WebVoyager drops from 77.72 to 66.31 and DeepSWE from 24.65 to 18.2. The card doesn't say what [offline] means, and we haven't found the harness that defines it. So we can't say whether web-navigation agents are the weak spot or whether that setup is just noisy. What we can say is that the Spark build and the reference build aren't interchangeable for agent work, and the card shows it.
NVIDIA's post says Super VL is "5x faster than the next best" open agentic model. Press coverage ties that to roughly 625 output tokens per second on four Blackwell B200s. That's a data center number. None of it describes a single Spark, and the card gives no Spark throughput at all. Nor did we find the leaderboard comparison behind it, since Artificial Analysis hadn't listed the model in the coverage we read. Read the claim as a vendor figure for a rack.
Our take: the useful part is that a 120B-class vision model now has an official, ungated recipe for one small machine. That's new. The part to be careful with is treating the card's headline numbers as what you'll get on it. We'd run your own task set on the NVFP4 build before choosing it over a smaller model. The row for it is in our open-weight models tracker, and we looked at the smaller sibling's 24GB label in our Bolt piece.
FAQ
Can Nemotron 3.5 Super VL run on one DGX Spark?
According to the NVFP4 model card, yes, using vLLM or SGLang with a 1M-token context. The BF16 build needs 4 H100s or 2 GB200s. We haven't run it ourselves.
How big is the NVFP4 checkpoint?
The files in the Hugging Face repo add up to about 75GB, by the repo's file list. Runtime memory will be higher once the KV cache and activations are counted.
Does the Spark setup support video input?
The card is inconsistent. Its deployment table lists text and image for the Spark and video for the larger setups, while the Spark quick-start recipe includes video options. Test it before relying on it.
What license covers the weights?
OpenMDW-1.1, per both model cards. Hugging Face shows the license tag as "other" because the license is named in the card rather than in the tag list.
Sources: Hugging Face, NVFP4 model card; Hugging Face, BF16 model card; NVIDIA, Nemotron 3.5 launch post; AI Weekly, Nvidia opens 120B Nemotron 3.5 Super VL.
