NVIDIA's Super VL: 28–72 Percent on Medical Benchmarks After Post-Training – Without Independent Verification
On October 15, 2026, NVIDIA released the open, multimodal agent model Nemotron 3.5 Super VL alongside two Bolt models. The company's benchmark chart reportedly shows that specialized post-trained versions beat Claude Opus 5 and GPT-6.1 Sol on five narrow tasks — but the base model trails the leading closed models, and none of the numbers have been independently verified.
What Was Released
The launch consists of three open models:
- Nemotron 3.5 Super VL – an open multimodal model with 120 billion parameters that, according to NVIDIA, can work across PDFs, diagrams, tables, images and web pages while following instructions and using tools. Available now.
- Bolt 74B-A9B (24 GB) – for local tasks. Available now.
- Bolt E4B (4 GB) – expected October 23 (Blockchain.News, citing NVIDIA's blog post of October 15, 2026).
The models are designed to run agents on everything from data center systems to desktop machines, and Super VL is, according to NVIDIA, intended to power agents working on tasks that may run for days or weeks. The company also states that enterprises can post-train the model on their own data to tailor it to their needs (AOL/Yahoo Finance).
How It Works: MoE and Open Weights
Super VL uses a Mixture-of-Experts architecture in which only 12 of the 120 billion parameters are active per token, according to NVIDIA's technical blog post (The Decoder). That makes inference cheap relative to the model's total size — a property well suited to high-volume agent workloads.
The model was developed with the Nemotron Coalition, a group that includes Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs.
The combination of open weights and lower inference cost is the point for enterprise use: NVIDIA markets the model with benefits such as lower cost and improved privacy, and highlights that enterprises can post-train it on their own data. This is precisely the path — base model plus domain-specific post-training — that the benchmark claims concern.
The Post-Training Claims: The Numbers NVIDIA Presents
In a chart reviewed by VentureBeat, NVIDIA claims that post-trained specialized versions of Super VL beat named Claude and GPT models on five narrow evaluations, with improvements of between 9 and 44 percent over the base model (VentureBeat):
- Medical claims: Accuracy rises, according to the chart, from roughly 28 to 72 percent after post-training, against around 38 percent for Claude Opus 5 on the same evaluation. ZDNET likewise reports a 44 percent lift on medical insurance claims, which according to NVIDIA places the model well above Claude Opus 5 (ZDNET).
- Resistance to security attacks: roughly 35 to 76 percent after training, against approximately 66 percent for GPT-6.1 Sol (VentureBeat, citing NVIDIA's chart).
Note that VentureBeat's figures are read from a chart and described as approximate. A partner press release from Mindbeam presents a parallel result: Post-trained Super VL for CUDA kernel generation reportedly raised KernelBench Level 2 from 22.8 to 82.1 percent, above GPT 5.6 Sol's 57 percent (Mindbeam/PRNewswire).
Note that Mindbeam lists "GPT 5.6 Sol," while VentureBeat lists "GPT-6.1 Sol" for a different comparison. This may involve different benchmarks and possibly different comparison models, but the naming discrepancy cannot be resolved with the available documentation.
The Important Caveats
So the story is "post-training that competes," not "open beats closed." Two facts underscore this:
The base model lags behind. In the overall Artificial Analysis ranking, Super VL sits well below the leading closed models, though it does beat NVIDIA's own Nemotron 3 Ultra (The Decoder). The claims about beating Claude and GPT therefore apply only to specialized, post-trained versions on narrow tasks — not to the base model in general.
No documentation or verification. NVIDIA has not provided task definitions, sample sizes or a test protocol for the five evaluations (VentureBeat). Blockchain.News also notes that the launch included no independent third-party verification of the performance claims. All the numbers — including the claim that Super VL is 5 times faster than the next best in its class, measured by Artificial Analysis against Ling 3.0 Flash and DeepSeek V4.1 Flash — are NVIDIA's own or its partners' claims: "As the fastest open agent model in its class — 5 times faster than the next best — it helps enterprise agents complete high-volume tasks faster, across workflows spanning days or weeks," NVIDIA writes in the release (as reproduced by ZDNET).
The Ecosystem Around the Launch
NVIDIA states that companies including Scale AI, AfterQuery, AI21, Alice, Mercor, ModelScope, Ramp, Dream, Lila Sciences and R1 are post-training or adapting Super VL for use cases ranging from finance and cybersecurity to scientific discovery and analysis of medical claims. In addition, platform integrations are listed with, among others, Atlassian, CrowdStrike, Perplexity and Salesforce (Blockchain.News).
The spread of partner results on launch day — Mindbeam is one example — is part of the narrative: NVIDIA is showcasing not just a model, but a pattern in which domain players build specialized versions and publish the numbers themselves.
What It Means – and What Remains
If the numbers hold up, they support a specific and practical use case: that open weights are a good starting point for cheap, private and domain-specific agent work, even when the base model never reaches the top of the general rankings. That is a different and more modest claim than open models beating the closed frontier models — and it is the only claim the evidence actually supports.
At the same time, readers cannot assess the results without protocols: we do not know how the five tasks are defined, how large the sample sizes are, or how the comparisons with Claude and GPT were conducted. The approximate chart figures cannot be confirmed without NVIDIA's primary sources.
Natural follow-up points are independent third-party verification of the five evaluations, the release of the 4 GB Bolt E4B package on October 23, and whether the naming discrepancy between "GPT-6.1 Sol" and "GPT 5.6 Sol" in the sources can be clarified.