Nvidia Adds Multimodal and Edge Models to Nemotron — Without Benchmark Details for Its Speed Claim

Nvidia says its new Nemotron 3.5 Super VL model has comparable power to its largest reasoning models at one-fifth the size, and that it produces outputs nearly five times faster than other models in the same class. SiliconANGLE reported the announcement on October 15, 2026. Its report names no benchmark, baseline models, hardware or test conditions, so the speed claim is a starting point for evaluation, not a result buyers can act on yet. ¹

What's in the Nvidia Nemotron lineup

The announcement covers three models for different deployment settings: a large multimodal model for agent tasks, and two smaller Bolt variants for local and edge use. ¹

• Nemotron 3.5 Super VL: an open, 120-billion-parameter model built with the Nemotron Coalition for multimodal agent tasks, including instruction following, tool use, multistep business workflows and visual understanding. ¹
• Bolt 74B-A9B 24GB: derived from Super VL, using a hybrid mixture-of-experts design with 74 billion total parameters, 9 billion active at a time, and two-bit quantization. Nvidia says it is designed to fit within 24GB of memory on an enterprise workstation or PC GPU, with writing and code debugging as example tasks. ¹
• Bolt E4B 4GB: a planned smaller model for edge deployment on GeForce RTX, Jetson and RTX/GTX Spark hardware. Nvidia lists possible uses including wearables, household devices, point-of-service systems, car dashboards and device assistants. The report says it is expected later that month. ¹

One wrinkle worth flagging before integration: the report's headline calls the large model “Nemotron 3.5 Super LV,” while its body calls it “Super VL.” This article follows the body text; teams should confirm the official model identifier before integration. ¹

Nvidia said it trained Super VL on about 26 trillion tokens across pretraining and continued pretraining, added 3.8 million post-training samples, and used reinforcement learning across 100 open environments. It also said it is releasing open datasets on Hugging Face covering agentic tool use, finance, instruction following, long context, safety, fairness and vision-language understanding. The report does not specify access or licensing terms. ¹

Nvidia provides the weights for post-training on an organization's own data and deployment on its own infrastructure. That gives teams the option to customize and host the model themselves, but the report does not describe the engineering effort involved or how a tuned model performs on a particular workflow. ¹

Nemotron 3.5 Super VL and Bolt compared: three models, one unanswered question

The table reflects what SiliconANGLE reported. These are design details and stated target uses, not performance results.

Model	Reported design	Stated target	What still needs testing
Nemotron 3.5 Super VL	Open, multimodal, 120B parameters 1	Tool use, multistep workflows and visual understanding 1	Task quality, latency and resource use on target hardware
Bolt 74B-A9B 24GB	Mixture-of-experts, 74B total / 9B active, two-bit quantization; designed to fit within 24GB 1	Local workstation agents; writing and code debugging 1	Quality and speed under actual workloads; memory and power behavior
Bolt E4B 4GB	Planned edge model 1	GeForce RTX, Jetson and RTX/GTX Spark; wearables, dashboards and device assistants 1	Release status, supported configurations and on-device performance

Parameter counts, quantization schemes and memory claims describe design choices, not how well a model will work. A mixture-of-experts architecture and two-bit quantization may inform engineering trade-offs, but do not establish tool-call accuracy or task completion rates.

The efficiency claim doesn't hold up to scrutiny yet

Nvidia's two headline numbers, one-fifth the size of its largest reasoning models and nearly five times the speed of other models in the same class, come without a named benchmark, baseline models, hardware specifications, batch size, workload description or definition of “faster” in SiliconANGLE's report. ¹

That leaves buyers without the details needed to assess the claim. Five times faster at what: time to first token or sustained throughput? On what hardware and at what batch size, against which models? The report does not say. It also provides no figures for agent success rates, tool-call accuracy, visual-task quality, power consumption, thermal behavior, failure recovery or cost per completed task. ¹

Nvidia's claims are a hypothesis to test, not a purchasing benchmark. Until a named benchmark is published with disclosed conditions, “nearly five times faster” is a claim to verify, not a figure to plan around. ¹

How to evaluate AI agent models: what buyers should measure

Run representative workflows on Super VL, the relevant Bolt variant and the alternatives your team already uses; a shared model catalog can make side-by-side comparison easier. Where relevant, test the task types Nvidia names: tool use, multistep business workflows, visual understanding, writing and code debugging. ¹

Keep these measures separate rather than collapsing them into a single “efficiency” number:

• Quality. Measure task completion, tool-call accuracy, visual interpretation and recovery from failed steps. The report supplies no results for these measures. ¹
• Latency and throughput. Record time to first token and full-response time, plus sustained throughput under your concurrency and hardware conditions. The report's speed claim has no stated benchmark or baseline. ¹
• Operating cost. Account for hardware, hosting, power, maintenance and serving costs. The report provides no prices or cost figures, so local deployment should not be assumed to be cheaper. ¹
• Deployment fit. Measure memory behavior under load, quantization accuracy loss, GPU compatibility and update process. Nvidia names memory figures and hardware families but the report does not give full deployment requirements. ¹
• Governance. Decide which workloads must run locally, which data each model can access, and who monitors quality and spend. The report does not specify compliance certifications, data-residency terms, uptime or an SLA for these models. ¹

An always-on agent is an operating requirement, not a model category. Test response-time consistency, power draw, update cadence and failure handling in the target environment; the announcement provides no measurements for those factors. ¹ The same discipline applies to any claim about long-running AI agents: a demonstration is not a reliability case.

What this means for edge AI models and architecture decisions

Super VL is presented as a broad multimodal agent model with open weights for post-training and self-hosting. The Bolt variants are presented for local or edge execution, with the 74B-A9B model described in terms of its memory fit and example tasks. That split offers a starting point for matching model size to deployment location, but it is not a verdict on suitability. ¹

For workstation deployments, check whether Bolt 74B-A9B 24GB's stated memory fit and example tasks match your GPU and response-time requirements. For embedded deployments, verify Bolt E4B 4GB's compatibility with the specific device; an edge focus does not establish suitability for every edge environment. ¹ A checklist of what to verify before on-device deployment is a useful template for that work, and the questions raised by local-model support in coding agents overlap with many of them.

Open weights give teams the option to post-train and host the model themselves. They do not establish how much engineering that will require or how the resulting model will perform against a hosted alternative. Weigh that option against measured quality and latency.

Next steps

Confirm the official model name, release status, access terms and deployment instructions before committing. The “Super LV” / “Super VL” discrepancy is reason to check the model identifier directly. Bolt E4B 4GB is described as launching later in the month in the October 15, 2026 report, so verify its availability before building around it. ¹

Build an evaluation matrix with separate fields for workload, quality, latency, memory, power, cost and deployment constraints. Record reproducible test details, including model version, quantization, hardware, context length, concurrency, workload mix and metric definitions; the announcement does not provide those test details. ¹

Start with the smallest model that clears your quality bar, then test whether moving up to Super VL improves task completion or reduces operational work. Do not assume either Bolt variant or Super VL wins by default. Before using Nvidia's datasets in production, check licensing and access terms; the report says Nvidia is releasing them but does not specify those terms. ¹

The practical question is whether Nvidia or an independent party publishes a named benchmark with disclosed baselines and hardware. Until then, treat “nearly five times faster” as a claim to verify, not a number to plan around. ¹

Frequently asked questions

What is Nemotron 3.5 Super VL?

It is described as an open 120-billion-parameter multimodal model for agent tasks, supporting instruction following, tool use, multistep business workflows and visual understanding. The source also says its weights can be used for post-training and deployment on an organization's infrastructure. ¹

What is Nemotron 3.5 Bolt 74B-A9B 24GB?

It is described as a smaller model based on Super VL, using a hybrid mixture-of-experts design with 9 billion active parameters out of 74 billion and two-bit quantization to fit within 24 gigabytes of memory. ¹

What is Bolt E4B 4GB designed for?

Nvidia described it as an edge model planned for GeForce RTX, Jetson and RTX/GTX Spark, with possible uses in wearables, household devices, point-of-service systems, dashboards and embedded interfaces. ¹

How fast is Nemotron 3.5 Super VL?

Nvidia said Super VL produces outputs nearly five times faster than other models in the same class, but the reported source provides no benchmark name, hardware, test conditions or underlying figures. ¹

Are Nemotron models cheaper to run than cloud models?

The source provides no prices or operating-cost measurements. Buyers should compare hardware, power, hosting, maintenance and model-serving costs rather than infer savings from the models' local or edge positioning. ¹

Sources

¹ “Nvidia expands Nemotron LLM family with hyper-efficient models for agentic systems,” SiliconANGLE AI, October 15, 2026. https://siliconangle.com/2026/10/15/nvidia-expands-nemotron-llm-family-with-hyper-efficient-models-for-agentic-systems/