• Bottom line: Use it as a pilot, not as a frontier-model replacement. The deciding number is 51.78 on SWE-bench Pro, as reported on Nvidia's model card.
• The model card lists the OpenMDW-1.1 license, up to 1 million tokens of context, and deployment on four H100 or two GB200 GPUs.
• The Decoder reports an Artificial Analysis Intelligence Index score of 26 against 58 for Claude Opus 5.5, so expect a clear gap to frontier closed models.
• Every accuracy and speed figure here is Nvidia-reported or relayed by The Decoder; we ran no hands-on tests and found no independent rerun.
• Token-count figures differ between Nvidia's card and The Decoder's report, so treat training totals as unsettled.

Our call on Nvidia's Nemotron 3.5 Super VL is Use it, as a pilot. Nvidia released the open 120-billion-parameter multimodal model on Oct. 15, 2026, and its model card reports 51.78 on SWE-bench Pro. The Decoder reports an Artificial Analysis Intelligence Index score of 26 against 58 for Claude Opus 5.5, so this is a capable open agent model, not a frontier substitute. This is a desk read of the model card and coverage, not hands-on testing.

What the model card actually says

The Hugging Face model card lists 120 billion total and 12 billion active parameters, context of up to 1 million tokens, and the OpenMDW-1.1 license, which the card describes as ready for commercial or non-commercial use. Deployment is documented on four H100 GPUs or two GB200 GPUs, with an NVFP4 release pointed to for DGX Spark. Nvidia's technical blog states neither the license nor the context length, so the card is the document to cite for both.

The number that decides the verdict

The deciding number is SWE-bench Pro at 51.78, from the model card, alongside 71.10 on SWE-bench Verified and 60.67 on Terminal-Bench 2.1. These are Nvidia-reported scores, not independent reruns. The Decoder separately reports an Artificial Analysis Intelligence Index v4.3.2 score of 26 against 58 for Claude Opus 5.5, which it says leads the index; we could not check that against Artificial Analysis directly. Together they place the model as a capable open agent model well below frontier closed models.

Speed claims come from Nvidia and favor its hardware

The speed claims are Nvidia's and are tied to its hardware. The Decoder reports about 10 hours for 10,000 SWE-bench Pro tasks on four B200 GPUs at roughly 51% accuracy, against 50 to 56 hours for two open competitors, read from a Nvidia chart. Nvidia's technical blog describes the result only as 4x faster completion of 10,000 coding tasks, and cites up to 5x single-GPU throughput over similar open models. Blackwell runs NVFP4 natively; older cards compute at 16 bits, per The Decoder.

What remains unverified

Several points remain unverified by us. We did not run the model, so none of the accuracy or throughput figures here is hands-on. The Decoder's adopter list, including Scale AI, Ramp, EY and Hebbia, and its cost projections come from Nvidia. Training totals also differ by page: the card lists a foundation corpus of more than 20 trillion tokens, The Decoder reports over 25 trillion including continued pre-training, and the technical blog gives no total. Treat token counts as unsettled.

Who should pilot it, and what to watch

Teams with document, chart or coding-agent workloads and existing H100 or Blackwell capacity can test it cheaply, since the weights and license are open. CharXiv reasoning at 80.82 on the card is the figure to reproduce for document work. Teams needing frontier-level reasoning should not expect parity given the Index gap. Watch for an independent Artificial Analysis listing, third-party SWE-bench Pro reruns, and throughput tests on non-Blackwell hardware.

The bottom line: Use it, as a pilot

Use it, but only as a bounded pilot. A 51.78 on SWE-bench Pro with an open license and a four-H100 footprint makes testing inexpensive, while a 26 on the Intelligence Index rules out treating it as a frontier replacement. Run it on your own task suite, measure accuracy and cost per task yourself, and commit hardware only if those results match Nvidia's reported figures.