NVIDIA releases Nemotron 3.5 Super VL with vision and a 1M-token context

Expand

Open original image ↗

NVIDIA has released Nemotron 3.5 Super VL, adding visual understanding to a model aimed at coding, research and agents that work through documents. It has 120 billion parameters, activates 12 billion per token and supports up to a million tokens of context. Text, images and video go in; text comes out. 1 2

That combination matters for tasks which mix formats. An agent reading a report may need to interpret a chart, find a figure in a table and then call a tool. NVIDIA’s pitch is that those steps can stay with one model, without handing the visual work to a separate system.

Where the benchmarks are strongest

NVIDIA’s model card reports 89.71 on GraphWalks for the BF16 version, alongside 89.52 on GPQA Diamond and 31.79 on text-only HLE. The visual results include 80.82 on CharXiv reasoning, 68.12 on English OCRBench v2 and 62.82 on MMLongBench-Doc. These are NVIDIA’s evaluations; Lumina has not independently reproduced them.

Expand

Open original image ↗

GraphWalks is a clear strength in the supplied comparison: the chart puts the larger Nemotron 3 Ultra at 63.3, against 89.7 for Super VL. The picture changes on other tasks. Qwen3.8 Flash-Next leads the chart on GPQA Diamond and HLE, and several models score higher on the coding tests.

The card gives Super VL 51.78 on SWE-bench Pro and 67.72 on SWE-bench Multilingual. On Pro, that puts it above Nemotron 3 Ultra’s 49.7 in the supplied chart, but below Qwen3.8 Flash-Next at 60.4. The smaller model improves on Ultra here without taking the lead.

A smaller deployment, with separate precision results

The model combines Mamba-2, attention and mixture-of-experts layers with a C-RADIO vision encoder. NVIDIA used distillation from Nemotron 3 Ultra and additional training for software engineering and visual tasks. Its multi-token prediction system can propose several output tokens at once for the model to check, reducing the work needed during generation. 1

Alongside BF16 weights, NVIDIA offers an NVFP4 version with lower memory requirements. The card lists configurations for one DGX Spark, one H200 or two H100s, among others. Context capacity depends on the deployment settings and available memory; the million-token maximum is shared by inputs, tool results and the response.

The two versions have separate benchmark columns. NVFP4 stays close on some tests—89.90 on GPQA Diamond versus 89.52 for BF16—but falls from 66.70 to 60.80 on Terminal-Bench 2.1 in the published table. A smaller checkpoint is useful, but it is worth checking the results for the version being deployed.

The weights are on Hugging Face under OpenMDW-1.1. NVIDIA is also releasing training data and recipes, giving teams material to adapt the model to their own documents and workflows.