Nvidia Launches Nemotron 3.5 With New Super VL and Bolt Models

Nvidia expanded its open Nemotron model family on Oct. 15, 2026, adding a 120-billion-parameter vision-language model called Nemotron 3.5 Super VL and two smaller, more efficient "Bolt" models built to run on local hardware instead of a data center. Super VL and the first Bolt model, Nemotron 3.5 Bolt 74B-A9B, are available now; a second, smaller Bolt model aimed at voice assistants and embedded devices arrives Oct. 23.

What Nvidia Announced

In a post titled "Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models", Nvidia VP of generative AI software Kari Briski laid out a now-familiar strategy for the company's open Nemotron line: ship one large, capable model for data-center-scale agentic work, then shrink it down into variants that run on a single GPU, a desktop workstation, or an embedded device. Super VL is the large end of that spectrum. The two Bolt models are the small end, designed to put similar agentic capability on an RTX PC, an Nvidia DGX Spark, or even a Jetson module without a cloud connection.

The announcement frames the three models as a single family rather than separate products, all trained and post-trained together through what Nvidia calls the Nemotron Coalition, its collaboration with outside partners on open data and training recipes.

It's also a deliberately timed release. Nvidia is pairing the software announcement with hardware: the company's RTX Spark laptops arrive Oct. 16, one day after Super VL and the first Bolt model go live, and Bolt 74B-A9B is explicitly called out as a model built to run on that hardware. The pitch to developers is that they don't have to choose between a capable multimodal agent and a model small enough to run offline — Nvidia wants the same Nemotron 3.5 lineage to cover both ends of that tradeoff, from a NIM-hosted endpoint in a data center to a laptop sitting on a desk.

What Is Nemotron 3.5 Super VL

Nemotron 3.5 Super VL is an open, 120-billion-parameter vision-language model built for agents that have to work with mixed content: text, images, charts, web pages and documents such as PDFs. Nvidia says the model matches the accuracy of its largest reasoning model on multimodal agentic tasks while using roughly one-fifth the parameter count, and it describes Super VL as "the fastest open agentic model in its class," claiming up to 5x the throughput of the next-best open model in the same category.

According to Nvidia's post, Super VL was pretrained on roughly 26 trillion tokens, including a continued pretraining phase, and then post-trained on about 3.8 million curated samples plus reinforcement learning across dozens of simulated environments. Nvidia highlights particular strength in optical character recognition, tables, charts, and parsing documents like web pages and PDFs — the kind of unglamorous but common inputs that trip up many multimodal models.

Architecturally, Nvidia's own technical documentation describes Super VL as pairing a Nemotron-H hybrid language backbone — which mixes Mamba-2 state-space layers with attention in a mixture-of-experts design — with a RADIO vision encoder, a vision projector, and a separate temporal video embedder reused from the earlier Nemotron Omni model family. That hybrid design is part of how Nvidia gets large-model accuracy out of a model that activates only a fraction of its total parameters per token.

For serving at scale, Nvidia is distributing Super VL through its NIM microservices, which the company says it refreshes roughly every two weeks, and it claims that running the model on its Dynamo inference software yields up to 8x more concurrent users and 85% lower latency compared with a baseline deployment. Nvidia lists Hugging Face, OpenRouter, ModelScope and NIM as the places developers can pull the model from today.

What Is Nemotron 3.5 Bolt

The Bolt side of the release is aimed squarely at hardware that sits on a desk rather than in a rack. Nemotron 3.5 Bolt 74B-A9B is a 74-billion-parameter mixture-of-experts model that activates about 9 billion parameters per token, and Nvidia has quantized it natively to 2 bits so the whole thing fits in 24GB of memory — within reach of a single high-end Nvidia RTX GPU or an Nvidia DGX Spark desktop system. It's available now on Nvidia's Hugging Face page, and Nvidia says it's already supported by common local-inference tools including llama.cpp, Ollama, LM Studio, Unsloth and vLLM. Nvidia also says Bolt 74B-A9B ranks as the #1 open model in its size class on the PinchBench agentic benchmark and the #2 open model on BrowseComp, and that it's built to run on the company's new RTX Spark laptops, which arrive Oct. 16.

A second, much smaller model — Nemotron 3.5 Bolt E4B 4GB — is scheduled to follow on Oct. 23. Nvidia positions it for latency-sensitive, always-on use cases like voice assistants, interactive characters and embedded interfaces, running on GeForce RTX laptops and desktops, Jetson edge modules, and RTX or DGX Spark hardware. Nvidia's post doesn't give a more detailed parameter breakdown for the E4B model beyond its name and memory footprint.

This kind of local, GPU-sized model also matters for anyone shopping for hardware to run it on — a decision that increasingly comes down to which GPU actually has enough memory and throughput for an agentic model instead of just a game benchmark.

How Nemotron 3.5 Compares to Earlier Nemotron Models

Nvidia's Nemotron line has historically shipped in Nano, Super and Ultra tiers, scaling from edge-friendly to data-center-scale. Super VL and the Bolt models slot into and extend that pattern — Super VL is a vision-language sibling to the text-focused Super and Ultra tiers, while Bolt pushes efficiency further down toward single-GPU and edge deployment than the standard Nano tier typically targets. Based on what Nvidia has published for this release and the architecture details in its own technical documentation, the lineup breaks down like this:

Nvidia's own benchmark claims for this release — the 5x throughput figure for Super VL, and the PinchBench and BrowseComp rankings for Bolt — are the company's internal comparisons against other open models in the same size class. Nvidia's post doesn't name every model in that comparison set, so those figures should be read as Nvidia's claims about its own release rather than independently audited results.

Licensing and Openness

Nvidia frames the entire Nemotron 3.5 release as open: open model weights, open training datasets, and published training recipes. The company says it is releasing the datasets used to train these models on Hugging Face, alongside the reinforcement-learning environments used in post-training through its NeMo Gym framework. That mirrors Nvidia's stated approach for the broader Nemotron family, which the company has consistently positioned as open enough for commercial use, fine-tuning and redistribution, in contrast to closed, API-only frontier models.

That open posture is also why Nemotron models keep showing up as a reference point in agent-tooling conversations — the same conversations that have been playing out around which coding assistant actually gets agentic work done, as in recent comparisons of GitHub Copilot against Claude Code. An open, benchmarked vision-language model that developers can self-host gives that ecosystem another building block to plug into agent frameworks rather than routing every call through a single vendor's API.

Who's Already Using Nemotron

Nvidia's announcement names a list of companies it says are already working with the Nemotron family for agentic and enterprise use cases, including Scale AI, EY, Hebbia, AI21, Mercor, Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys. Nvidia doesn't specify in the post which of these partners are using Super VL specifically versus other Nemotron models, or in what capacity, so treat the list as evidence of ecosystem traction rather than confirmation that each company has deployed Super VL or Bolt in production.

Where to Get Nemotron 3.5

Developers can pull Nemotron 3.5 Super VL from the Hugging Face model card, OpenRouter, ModelScope, or run it as a hosted Nvidia NIM microservice through Nvidia's Nemotron foundation model page. The Bolt 74B-A9B model is also on Hugging Face now, with day-one support across llama.cpp, Ollama, LM Studio, Unsloth and vLLM, meaning it can be run through most of the popular local-inference stacks without waiting on separate integration work. Nvidia's own technical documentation for the Megatron Bridge and NeMo AutoModel frameworks also lists checkpoint names and conversion recipes for teams that want to fine-tune or adapt the models rather than run them as-is.

How Super VL Fits Into the Open-Model Race

Super VL's release lands in the middle of a crowded field of frontier and open models that are increasingly compared head-to-head on agentic and multimodal tasks, from closed frontier systems to other open-weight releases. Pandromeda has covered that broader competitive landscape in pieces like a detailed look at how the current top closed models stack up against each other. Nvidia doesn't position Super VL as a direct replacement for those general-purpose assistants; instead, it's pitched specifically at developers building agents that need to parse documents, charts and screenshots as part of a larger workflow, and who want to self-host or fine-tune the model rather than call a closed API. That's a narrower, more infrastructure-oriented pitch than a consumer chat assistant, but it's also why Nvidia emphasizes open weights, published datasets and NeMo Gym training environments rather than a single leaderboard score.

What This Means

The Super VL and Bolt release continues a pattern Nvidia has leaned on since earlier Nemotron generations: rather than shipping a single flagship model, the company is building a tiered family so the same general capability — agentic reasoning over mixed text, image and document inputs — can run on anything from a DGX data-center rack down to a desktop GPU or an embedded module. For developers, the more immediate practical question is less about Nvidia's internal benchmark claims and more about fit: Super VL targets teams that need the strongest open multimodal agent and don't mind data-center-scale serving, while the Bolt models target teams that want similar agentic behavior running locally, privately and without a per-token API bill.

The next milestone to watch is the Oct. 23 launch of Bolt E4B, the smallest model in the family, which will be the real test of whether Nvidia's "desktop to data center" pitch extends all the way down to voice assistants and embedded hardware. Nvidia also says it will keep refreshing its NIM-hosted version of Super VL roughly every two weeks, so expect incremental capability and performance updates rather than a single static release.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage