NVIDIA released the open Nemotron 3.5 family on October 15, 2026: a 120 billion parameter multimodal model called Super VL, plus smaller Bolt models for local and edge use. The headline claim, as reported by VentureBeat, is that customized versions beat Anthropic's Claude Opus 5 and OpenAI's GPT-6.1 Sol on five specialized tasks. That claim is narrower than it sounds. This article lays out what was released, what the benchmark chart does and does not show, how the speed figure should be read, and what the license and enterprise terms imply.
NVIDIA
Anthropic
OpenAIWhat was released
Per VentureBeat's report, Nemotron 3.5 Super VL combines language and visual understanding and handles PDFs, charts, tables, images and webpages, along with instruction following and tool use. The Hugging Face listing names it 120B-A12B, which suggests about 12 billion active parameters, and AI Weekly's summary states a mixture-of-experts design with 12 billion active and a hybrid Mamba-Transformer backbone and 1 million token context. VentureBeat's own article does not state the active count, so treat that figure as reported by one secondary source.
Two smaller models were announced. Bolt 74B-A9B is a 74 billion parameter model with 9 billion active, using native 2-bit compression to fit in 24 GB of memory, aimed at local coding, web research and multistep tasks. Bolt E4B, a 4 GB model for conversational and on-device agents, is scheduled for October 23. Training used about 26 trillion tokens, 3.8 million post-training samples and reinforcement learning across roughly 100 open environments, and NVIDIA says training recipes and datasets are being released too.
How to read the benchmark claim
NVIDIA's chart covers five evaluations: document intelligence, security attack resilience, financial receipt extraction, scientific experimentation and medical insurance claims. It shows post-training gains of 9 percent to 44 percent over the base Super VL. Two data points were reported with enough detail to chart: medical claims rising from about 28 percent to about 72 percent against about 38 percent for Claude Opus 5, and security attack resilience rising from about 35 percent to about 76 percent against about 66 percent for GPT-6.1 Sol.
Customized small models versus frontier models, on one task each
Approximate scores read from NVIDIA's chart as reported by VentureBeat. Task definitions and sample sizes were not published.
The comparison is between a model customized for a task and general-purpose frontier models that were not. The base Super VL scored below both frontier models on these two tasks, so the result says more about the effect of task-specific training than about general capability. VentureBeat states this directly: the comparisons are for customized models on specific tasks, not general superiority. NVIDIA also has not released task definitions, sample sizes or testing protocols, which limits how far the numbers can be interpreted.
The base model trailed both frontier models. The gain came from post-training on narrow tasks, which is the real finding.
The speed figure
NVIDIA reports about 625 output tokens per second for Super VL at an Artificial Analysis Intelligence Index score of about 26, with 100,000 input tokens, measured on four B200 GPUs. It claims this is up to 5 times faster than a comparable open model, citing Inkling Small at about 125 tokens per second at a similar index score. Several caveats from the source report apply: the figure measures output generation, not end-to-end workflow time; on the composite index, 26 trails other open models such as GLM-5.3-Flash at 42 and Qwen3.8-Flash-Next at 40; and the index does not evaluate image inputs, so it does not capture the document-reading strength NVIDIA is emphasizing.
Why NVIDIA releases open models
NVIDIA sells GPUs, and open models that run well on them widen the market for that hardware. Super VL is optimized for Blackwell and Hopper systems, including DGX Spark, DGX Station and RTX PRO 6000 Blackwell, and can run on Ampere GPUs. A rate-limited NVIDIA-hosted API is free for development, while production deployment through NIM microservices carries an NVIDIA AI Enterprise license of $4,500 per GPU per year, excluding hardware, hosting and operations. The weights themselves use the OpenMDW license, which permits commercial use, fine-tuning and redistribution. For a deeper look at this strategic logic, see our analysis of Nemotron 4.
The launch partner list also shows where NVIDIA expects the models to be used: post-training customers include Scale AI (for Skydance and EY), AfterQuery (for Hebbia), AI21, Alice, Mercor, Ramp, Dream, Lila Sciences and R1, with Atlassian, CrowdStrike, Dassault Systemes, Perplexity, Salesforce and Synopsys listed as platform integrators.
What this means for model selection
The practical lesson is a familiar one: for a well-defined, repeated task with labeled data, a smaller open model trained on that task can match or exceed a general frontier model, at lower serving cost. The open question is the cost of getting there, since customization requires data, evaluation and engineering that a general API does not. For teams without that investment, a general frontier model remains the simpler default. Products that let users switch between model families, such as Metir AI, make it easier to test that trade-off on real workloads before committing to a custom model.
What remains unknown
- Task definitions and sample sizes. NVIDIA has not published them, so the five-task result cannot be reproduced.
- General performance. No comparison against frontier models on broad benchmarks was reported for the customized variants.
- Active parameter count. Reported as 12 billion by one summary but not in the primary report.
- Bolt benchmark scores. Bolt is described as top in its class on PinchBench and second on BrowseComp, but no scores or class definition were given.
- Total cost. The $4,500 per GPU per year license is only one component of production cost.
The takeaway
Nemotron 3.5 Super VL is an open multimodal model with permissive licensing and a strong throughput claim, and the benchmark chart illustrates how much task-specific post-training can add. It does not show that an open model has overtaken frontier models in general. The most useful interpretation is that customization can close or reverse gaps on narrow tasks, while the details needed to verify NVIDIA's numbers remain unpublished.
Sources:
- Nvidia's new open Nemotron 3.5 Super VL with post-training beats Claude Opus 5, GPT-6.1 Sol on specialized task benchmarks | VentureBeat
- AI News Today, October 15: Top Stories | AI Weekly
Image credits
Header and in-body image: NVIDIA headquarters in Santa Clara, California, photographed by Coolcaesar on 4 August 2018, via Wikimedia Commons, licensed CC BY-SA 4.0. Illustrative of the company only; it does not depict the Nemotron models or their training hardware.