iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
NVIDIA’s October 15, 2026 Nemotron 3.5 release adds a 120-billion-parameter multimodal model for long-running agents and two compact “Bolt” models aimed at local and edge hardware. Alongside the existing 30-billion-parameter Lightning model and the Switchyard routing layer, the expanded family covers enterprise workflows, high-volume specialist tasks, desktop GPUs and embedded devices.
What NVIDIA added on October 15, 2026
The headline release is Nemotron 3.5 Super VL, an open multimodal model intended to follow instructions, call tools and complete multistep workflows involving webpages, PDFs and visual information. NVIDIA positions it for “always-on” agents that may need to observe inputs, plan actions and continue working across a business process rather than answer a single chat prompt.
NVIDIA also introduced two efficiency-focused models under the Bolt name. Bolt 74B-A9B uses native 2-bit quantization to fit a 24GB memory footprint for local agentic work on RTX systems and DGX Spark. Bolt E4B has a 4GB footprint for fast conversational and physical-agent applications on GeForce RTX, Jetson and RTX/DGX Spark hardware; NVIDIA listed October 23, 2026 as its scheduled availability date.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsThese releases extend, rather than replace, the rest of the Nemotron stack. Lightning is designed for specialized, high-volume jobs, while Switchyard is an open-source router that chooses a model according to quality, latency and cost priorities.
#1 Best Overall
- AI Performance: 767 AI TOPS
- OC mode: 2632 MHz (OC mode)/ 2602 MHz (Default mode)
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Axial-tech fan design features a smaller fan hub that facilitates longer blades and a barrier ring that increases downward air pressure
- A 2.5-slot design maximizes compatibility and cooling efficiency for superior performance in small chassis
How the Nemotron models differ
| Model or component | Primary role | Published size or memory | Deployment emphasis |
|---|---|---|---|
| Nemotron 3.5 Super VL | Multimodal instruction following, tool use, visual understanding and multistep business workflows | 120 billion parameters | Enterprise servers, data centers and cloud; NVIDIA also lists RTX, DGX and RTX PRO platforms in its deployment coverage |
| Nemotron 3.5 Bolt 74B-A9B | Local web research, writing and code-debugging agents | 74B-A9B; 24GB footprint using native 2-bit quantization | RTX PCs and workstations, DGX Spark and other local systems with at least the stated memory capacity |
| Nemotron 3.5 Bolt E4B | Fast conversational and physical-agent workloads, including voice assistants and embedded interfaces | 4GB footprint; parameter count not stated | GeForce RTX, Jetson and RTX/DGX Spark; scheduled by NVIDIA for October 23, 2026 |
| Nemotron 3.5 Lightning | Specialized, high-volume tasks inside larger multi-agent systems | 30 billion parameters | High-throughput serving; exact hardware requirement not stated |
| Switchyard | Routes requests to the model that best fits quality, latency and cost goals | Router, not a language model | Open-source orchestration across a multi-model deployment |
Super VL is the capability and modality choice when an agent must understand images or documents and use tools. Lightning targets repeated specialist calls where throughput matters most. Bolt trades broad model scale for the ability to keep intelligence close to the user, device or private network.
Which Nemotron model should you use?
Choose Super VL for document- and tool-heavy enterprise agents
Use Super VL when the workflow combines visual inputs with planning and actions: for example, an agent that reads a PDF, checks a webpage, extracts information and then invokes business tools. NVIDIA describes it as comparable with its largest reasoning model at about one-fifth the parameter count, but that is a vendor positioning claim rather than an independent evaluation.
Choose Bolt 74B-A9B for local agents on a 24GB-class system
Bolt 74B-A9B is the practical starting point for an RTX desktop or workstation when data should remain on the machine. Its 24GB footprint is the published target, not a guarantee that every application will have 24GB free: the operating system, runtime, context length and other models also consume memory. NVIDIA lists llama.cpp, Ollama, LM Studio, Unsloth and vLLM support, giving developers several local serving paths.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Rank #2
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5060
- Integrated with 8GB GDDR7 128bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
Choose Bolt E4B for constrained and embedded devices
Bolt E4B is aimed at low-latency interactions where a 4GB footprint matters more than maximum reasoning depth. The stated examples include local voice assistants, interactive characters and embedded interfaces running on GeForce RTX or Jetson-class hardware.
Use Lightning for repeated specialist calls
Lightning fits a larger agent system as a fast worker for a narrow task, such as classification, extraction or another high-volume operation. It is not presented as the family’s general multimodal flagship.
Add Switchyard when a single model is wasteful
Switchyard can send difficult requests to a stronger model, routine requests to a faster one and cost-sensitive requests to a smaller one. That routing strategy is intended to make a multi-agent service more economical without forcing every request through the largest model.
Rank #3
- Powered by the NVIDIA Blackwell architecture and DLSS 4
- Powered by GeForce RTX 5070 Ti
- Integrated with 16GB GDDR7 256bit memory interface
- PCIe 5.0
- WINDFORCE cooling system
Where Nemotron can run
NVIDIA’s deployment story spans RTX PCs and workstations, DGX Spark, DGX Station, Jetson, RTX PRO systems, conventional data centers and cloud environments. Super VL is distributed through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM. Bolt 74B-A9B is listed for common local runtimes, including llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
For an on-premises or offline design, the important distinction is between the model and the serving stack. Open weights let an organization place inference on infrastructure it controls, but an individual package may still need a one-time download, a compatible runtime, model files and any separately configured tools. NVIDIA’s announcement does not establish that every NIM or community runtime works without network access.
What “open” means for customization and control
NVIDIA’s value proposition is control over deployment and post-training. For Super VL, NVIDIA identifies open weights, datasets and training recipes, allowing organizations to adapt the model on their own data and run it on their own infrastructure. That can support privacy, data-residency and governance requirements that are harder to meet when inference is tied to a hosted endpoint. Teams still need to check the applicable model license, hardware capacity and the security of any tools connected to an agent.
Rank #4
- Powered by the NVIDIA Blackwell architecture and DLSS 4 OC mode: 2640MHz/Default mode: 2610MHz (Boost Clock)
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.125-slot design with massive fin array optimized for airflow from three Axial-tech fans
- Phase-change GPU thermal pad helps ensure optimal thermal performance and longevity, outlasting traditional thermal paste for graphics cards under heavy loads
How strong are NVIDIA’s speed and accuracy claims?
The comparative figures are NVIDIA-reported results, not neutral industry benchmarks. NVIDIA claims Super VL can deliver up to five times the throughput of the next-best model in its class. For Lightning, it reports up to four times faster output and 30% faster agentic task completion than other models in its class. NVIDIA’s internal Switchyard benchmark says routing reduced task-completion cost to nearly one-third of using Opus 4.8 alone while maintaining frontier-level accuracy.
An earlier 2025 Llama Nemotron release carried NVIDIA claims of up to a 20% accuracy improvement and five-times faster inference after post-training. Those numbers, like the 2026 comparisons, depend on the chosen models, prompts, hardware, quantization, concurrency and evaluation tasks. The official material available for this release does not provide independent validation sufficient to treat them as universal performance guarantees.
Quick Recap
Availability and practical next steps
- Match the workload to the model. Decide whether you need multimodal reasoning, high-volume specialist calls or local inference.
- Check memory before downloading. Treat the published 24GB and 4GB figures as model-footprint targets and reserve additional capacity for the runtime and context.
- Select a serving path. Super VL is offered through NVIDIA NIM and listed model platforms; Bolt 74B-A9B has support across several local runtimes.
- Test with your own tasks. Measure latency, throughput, tool-call reliability and accuracy on representative documents and prompts rather than relying only on vendor comparisons.
- Plan routing and safeguards. If several models are deployed, use a router such as Switchyard and enforce tool permissions, logging and data-retention rules.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
