Nvidia says its new open vision-language model beats Claude Opus 5 and GPT-6.1 Sol on specialized benchmarks after post-training, though independent tests have yet to arrive
Nvidia has a new model, and it is picking fights with the biggest names in AI.
The company introduced Nemotron 3.5 Super VL, a vision-language model that, according to VentureBeat’s reporting, outperforms Anthropic’s Claude Opus 5 and OpenAI’s GPT-6.1 Sol on specialized benchmarks after post-training.
The catch is that no independent benchmarking has been published yet.
What Nvidia actually built #
The full name is Nemotron 3.5 Super VL 120B A12B, which reads like a license plate but carries real information. The model was officially documented on October 15, 2026.
It packs around 121 billion total parameters and a context window of 1 million tokens. A context window is the model’s working memory: how much text it can hold in view at once.
Under the hood, the model pairs two components. The language side runs on a hybrid mixture-of-experts backbone called NemotronH_Omni_Reasoning_V3. The vision side uses a RADIO v2.5-H encoder, which lets the model take in images alongside text.
In this model, there are 512 routed experts, and 22 of them activate for each token. The rest sit idle, which is the point: you get the knowledge of a large model while paying the compute bill of a much smaller one.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Nvidia also layered in Mamba-Transformer efficiencies and multi-token prediction. Multi-token prediction lets the model guess several words ahead instead of one at a time, which can speed up output.
A model built for agents, not chit-chat #
Nvidia is not pitching this as a general-purpose assistant. The model primarily targets efficiency in agentic tasks.
The benchmark claims come with a specific qualifier: the model’s edge over Claude Opus 5 and GPT-6.1 Sol shows up on specialized benchmarks, after post-training.
The Nemotron track record #
The original Nemotron 3 family launched in December 2025, with an emphasis on open weights and permissive licenses. A Super model followed in March 2026, and a Lightning model arrived in August 2026.
Earlier models in the series reportedly delivered up to 2.2 times the throughput of GPT-OSS-120B while holding competitive accuracy across various benchmarks.
The open-weights approach is central to Nvidia’s strategy. Open weights mean companies can download the model, run it on their own infrastructure, and customize it for their needs. Closed models from Anthropic and OpenAI, by contrast, are typically accessed through paid APIs.
What this means #
The most immediate question is whether the benchmark claims hold up. Without independent verification, the comparison to Claude Opus 5 and GPT-6.1 Sol is a company claim, not a settled fact.
The next checkpoint is GTC Berlin, scheduled for October 20-22, 2026, where the model is set to be discussed further. Watch for details on availability, licensing terms, and whether Nvidia shares the underlying benchmark methodology.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our