Nvidia Releases Nemotron 3.5 Super VL, an Open Multimodal Model for Enterprise Agents

Nvidia on October 15, 2026 released Nemotron 3.5 Super VL, an open weight model with 120 billion parameters that reads text, images, documents and video and is built to run enterprise agents. It shipped together with Nemotron 3.5 Bolt 74B-A9B 24GB, a smaller model compressed to fit in 24GB of memory for local agents, while a third model, Bolt E4B 4GB, follows on October 23, the company said.

The release extends Nvidia's line of open weight models after Nemotron 3 Ultra on June 4, 2026 and the text model Nemotron 3.5 Lightning on August 11, 2026. Super VL was distilled from Ultra, Nvidia says, and Artificial Analysis now scores it higher than Ultra, although its 120 billion parameters are about a fifth of Ultra's 550 billion.

What did Nvidia release?

Nvidia published two open weight models on Hugging Face on October 15, 2026, both under the OpenMDW 1.1 license: Nemotron 3.5 Super VL for data centers and workstations, and Nemotron 3.5 Bolt 74B-A9B 24GB for local PCs.

Super VL is a mixture of experts model, which means only part of the network runs for each token: 12 billion of its 120 billion parameters are active, according to the model card. It combines Mamba-2, expert and attention layers with a vision encoder and supports up to 1 million tokens of context. Nvidia says it trained the model on about 26 trillion tokens and post-trained it with roughly 3.8 million samples and reinforcement learning. The main download uses NVFP4, a 4-bit number format for Nvidia's Blackwell GPUs; Artificial Analysis found it within 0.1 points of the full precision weights on its index.

Bolt 74B-A9B has 74 billion total and 9 billion active parameters, uses native 2-bit quantization to fit in 24GB, and handles up to 256,000 tokens of context, per its model card. Nvidia says it is built for RTX Spark laptops, which go on sale October 16, and runs in llama.cpp, Ollama, LM Studio and vLLM. When Microsoft opened RTX Spark preorders on October 7, it named an upcoming Nvidia Nemotron model with more than 70 billion parameters that fits in just over 20GB among the first models tuned for those PCs; Bolt's published specs match that description.

How does Super VL score in independent tests?

Artificial Analysis, which runs its own evaluations of AI models, scored Nemotron 3.5 Super VL at 26 on its Intelligence Index on October 15, 2026, the highest score of any open weight model from a US developer it has measured. That puts it just ahead of Thinking Machines Lab's Inkling Small (25.7) and ahead of Nemotron 3 Ultra (23).

The score is twice that of its direct predecessor, Nemotron 3 Super, and Super VL uses about 60% fewer output tokens per task, according to Artificial Analysis. US open models still trail the best from the United Arab Emirates (K2 Horizon, 31) and China (MiMo-V2.6-Pro, 46) on the same index.

Speed is where Nvidia's headline claim comes from. Artificial Analysis measured about 630 output tokens per second with 100,000 token prompts on a pre-release DeepInfra endpoint, roughly five times Inkling Small's 124. Nvidia's own speed chart is built on the same index and names Inkling Small as the model it is "5x faster" than, so the claim compares Super VL with one similarly scored open model, not with closed models.

Bolt 74B-A9B ranks fourth among models under 100 billion parameters on the index and first among US models in that size class, Artificial Analysis says.

Does Super VL beat Claude and GPT-6 models?

Only after customization, and only in Nvidia's own chart of five specialized workflows. In Nvidia's technical blog, versions of Super VL that partners post-trained on their own data score above a named closed model in each of the five tasks, with accuracy measured by those organizations inside their workflows.

The rivals differ per task: Claude Opus 5 for document intelligence (Scale AI for Skydance) and medical insurance claims (R1), OpenAI's GPT-6.1 Sol for security attack resilience (AI21 and Alice), GPT-6 Luna for financial accounting (Ramp) and GPT-6 Astra for scientific experimentation (Lila Sciences). In the same chart, the out-of-the-box Super VL scores below the rival in all five tasks; post-training adds between 9% and 44%, by Nvidia's labels. VentureBeat reads the medical claims bars as about 28% before and 72% after post-training, against about 38% for Claude Opus 5.

The chart has no exact scores, and the technical blog gives no task definitions or sample sizes. Two of the five comparisons use Claude Opus 5, although Anthropic released Claude Opus 5.5 on September 22, 2026. GPT-6.1 Sol is OpenAI's current Sol model, with a system card dated September 29, 2026.

Who can use it, and what does it cost?

Super VL is available now on Hugging Face, OpenRouter, ModelScope and as an Nvidia NIM microservice, and the weights are free to download, Nvidia said on October 15, 2026.

On OpenRouter, the lowest listed price on October 15 was $0.15 per million input tokens and $0.50 per million output tokens through Together, and a rate limited free variant is also listed. Nvidia quotes $4,500 per GPU per year for an NVIDIA AI Enterprise license for production use of NIM, according to VentureBeat. Nvidia names Atlassian, CrowdStrike, Perplexity, Salesforce and Dassault Systèmes among the software companies integrating the model.

What we don't know yet

• The exact scores, task definitions and sample sizes behind the five specialized comparisons, and how a post-trained Super VL fares against Claude Opus 5.5.
• How fast Bolt 74B-A9B runs on the RTX Spark laptops that go on sale October 16; Artificial Analysis gives no speed figure for it.
• How Bolt E4B 4GB performs, since it is not due until October 23.