NVIDIA’s October 15, 2026 Nemotron 3.5 expansion adds a 120-billion-parameter multimodal model for long-running, tool-using agents and two smaller models designed to run locally or at the edge. Super VL targets enterprise workflows, Lightning targets specialized high-volume tasks, and Bolt targets useful agent intelligence within desktop and embedded memory limits.
What NVIDIA added to Nemotron
The release broadens Nemotron from a model family aimed mainly at server deployments into a portfolio spanning data centers, workstations, gaming PCs and embedded devices. NVIDIA announced Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B on October 15, 2026. Super VL and Bolt 74B-A9B were available at announcement; NVIDIA said Bolt E4B was scheduled for October 23.
As an Amazon Associate I earn from qualifying purchases.
The lineup is intended to let a developer assign different parts of an agent system to models with different capability, latency, memory and cost profiles rather than sending every request to the largest model.
How Super VL, Lightning and Bolt differ
| Model | Primary role | Scale or memory claim | Typical workloads | Deployment emphasis |
|---|---|---|---|---|
| Nemotron 3.5 Super VL | Multimodal instruction following and complex agent workflows | 120 billion parameters | Tool use, multistep business processes, webpage and PDF analysis, visual understanding | Enterprise servers, data centers and cloud; also supported across NVIDIA’s broader infrastructure |
| Nemotron 3.5 Lightning | Specialized, high-volume work inside multi-agent systems | 30-billion-parameter mixture of experts | Repeated domain tasks where output speed and throughput matter | Used as a fast specialist alongside other models |
| Nemotron 3.5 Bolt 74B-A9B | Local agentic intelligence | 24GB memory footprint using native 2-bit quantization | Web research, writing and code debugging | Local RTX systems and DGX Spark |
| Nemotron 3.5 Bolt E4B | Fast conversational and physical-agent interactions | 4GB memory footprint | Voice assistants, interactive characters and embedded interfaces | GeForce RTX, Jetson and RTX or DGX Spark systems |
The “A9B” designation appears in Bolt 74B-A9B’s name, while NVIDIA’s announcement presents the model as a 24GB local deployment option. The announcement does not provide a separate independent measurement of its effective active-parameter count, so the designation should not be treated as a benchmark result.
#1 Best Overall
- ATTENTION BUYERS: EXPORTING this item outside the U.S. requires strict compliance with the Bureau of Industry and Security (BIS) U.S. Export Administration Regulations (EAR). Non-compliance carries fines up to $1 MILLION PER VIOLATION AND UP TO 20 YEARS IMPRISONMENT for both the buyer and seller. To uphold our regulatory obligations, we strictly follow BIS "Know Your Customer" guidance by screening each order and immediately aborting transactions showing unresolved regulatory red flags. ***
- [Double-Flow-Through Design] The RTX PRO 6000 Blackwell features a double-flow-through cooling design, optimizing efficiency and airflow to sustain peak performance under 600W power loads. | [5th Gen Tensor Cores] Deliver up to 3X the performance of the previous generation and support for FP4 precision for faster AI model processing times with reduced memory usage, enabling local fine-tuning of LLMs and generative AI | [4th Gen Ray Tracing Cores] Double the ray-triangle intersection rate of the previous generation to create photoreal, physically accurate scenes and immersive 3D designs with RTX Mega Geometry, which enables up to 100X more ray-traced triangles.
- [PCIe Gen 5] Support for PCIe Gen 5 provides double the bandwidth of PCIe Gen 4, improving data-transfer speeds from CPU memory and unlocking faster performance for data-intensive tasks like AI, data science, and 3D modeling. | [GDDR7 Memory] With 96 GB of GPU memory and 1.8 TB ps bandwidth, it can tackle massive 3D and AI projects, fine-tune AI models locally, explore large-scale VR environments, and drive larger multi-app workflows.
- [DisplayPort 2.1] Achieve unparalleled visual clarity and performance, driving high resolution displays at up to 8K at 240 Hz and 16K at 60 Hz. Increased bandwidth enables seamless multi-monitor setups while HDR and higher color depth support ensures superior color accuracy for precision work, such as video editing, 3D design, and live broadcasting.
- [NVIDIA Blackwell Streaming Multiprocessor] The new SM features increased processing throughput, and new neural shaders that integrate neural networks inside of programmable shaders | DLSS 4: Multi Frame Generation ensures ultra-smooth frame pacing for lifelike simulations. | [Universal MIG] Divide a single RTX PRO 6000 Blackwell into multiple isolated instances, each with dedicated resources, allowing for concurrent execution of multiple workloads, optimized GPU utilization, and secure isolation of different applications or users. [WARRANTY] 3 YR Manufacturer's Warranty. Bulk OEM Packaging. Retail Packaging is NOT included.
Super VL: the multimodal enterprise model
Super VL is built for agents that must interpret images and documents, follow instructions, call tools and carry out several steps without stopping after a single chat response. NVIDIA describes it as comparable with its largest reasoning model at roughly one-fifth the parameter count and claims up to five-times faster throughput than the next-best model in its class. Those are NVIDIA comparisons, not independently verified results.
Lightning: a fast specialist
Lightning is a 30-billion-parameter mixture-of-experts model intended for specialized, high-volume work within a larger agent architecture. NVIDIA reports up to four-times faster output and 30% faster agentic task completion than other models in its class. A sensible use is to route repetitive, well-defined subtasks to Lightning while reserving a larger model for ambiguous reasoning or rich multimodal input.
Bolt: intelligence per gigabyte
Bolt prioritizes local execution. The 74B-A9B version uses native 2-bit quantization to fit a 24GB memory footprint, making it a candidate for an RTX PC, workstation or DGX Spark when a project needs local web research, writing or debugging. Bolt E4B reduces the stated footprint to 4GB and is aimed at responsive voice, character and embedded interfaces. NVIDIA listed E4B as forthcoming on October 23, rather than available on launch day.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Which Nemotron model should you use?
- Choose Super VL when the agent must understand mixed text and visual inputs, inspect webpages or PDFs, use tools across a multistep workflow, or support demanding business processes.
- Choose Lightning when the task is specialized and repeated at high volume, and lower latency or serving cost is more important than maximum generality.
- Choose Bolt 74B-A9B when data should stay on a local RTX or DGX Spark system and the available accelerator has about 24GB of usable memory for the model.
- Choose Bolt E4B for a small local or embedded conversational interface, subject to its stated October 23 availability and the runtime support available for the target device.
- Combine them when an application needs a router or multi-agent design: a fast specialist can handle routine work, while Super VL handles difficult multimodal steps.
NVIDIA’s open-source Switchyard router is designed for this kind of allocation. It directs requests to an appropriate model according to quality, latency and cost priorities. NVIDIA says an internal benchmark reduced task-completion cost to nearly one-third of using Opus 4.8 alone while maintaining frontier-level accuracy. That result is an NVIDIA internal measurement, not a neutral cross-vendor benchmark.
Rank #2
- NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
- 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
- PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
- NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
- Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads
Where Nemotron can run
NVIDIA says the models run across RTX PCs and workstations, DGX Spark, DGX Station, Jetson devices, RTX PRO systems, conventional data centers and cloud environments. The practical choice depends on memory, latency, privacy and how much operational control the team needs.
Local and edge deployment
Bolt 74B-A9B has announced support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. That gives developers several routes to local testing and serving, but a 24GB footprint is a model requirement rather than a guarantee that every 24GB graphics card will deliver the same speed. Runtime overhead, context length, quantization settings and other applications sharing memory can change the result.
Bolt E4B’s 4GB claim is better suited to constrained edge hardware, including supported GeForce RTX and Jetson deployments. NVIDIA also names DGX Spark as a target for both local experimentation and more capable workstation-style deployments.
Server and cloud deployment
Super VL is available through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM. NVIDIA’s NIM packaging is intended to simplify deployment in data-center environments, while the company’s earlier Llama Nemotron release described Nano, Super and Ultra NIM microservices and distribution through build.nvidia.com and Hugging Face. Availability through a platform does not by itself establish identical performance, pricing or service terms across providers.
Rank #3
- Professional GPU with Blackwell Architecture
- Blackwell Architecture
- 24GB GDDR7 with PCIe 5.0 & Ray Tracing
- AI Workstation
What “open” means for an agent project
NVIDIA’s pitch is control over where an agent runs and how it evolves. For Super VL, the company says open weights, datasets and training recipes are provided so organizations can post-train on their own data and deploy on their own infrastructure. That can help with data-residency requirements, private documents and domain-specific behavior.
Running an open model on premises or offline is an architecture decision, not an automatic property of the model. A team must obtain the weights and runtime, provide sufficient hardware, configure local storage and disable or restrict network access if isolation is required. Cloud-hosted wrappers may add their own telemetry, retention and availability policies, so those terms need separate review.
How strong are NVIDIA’s speed and accuracy claims?
The headline numbers are vendor-reported. For Super VL, NVIDIA claims up to five-times faster throughput than the next-best model in its class. For Lightning, it reports up to four-times faster output and 30% faster task completion. NVIDIA’s earlier 2025 Llama Nemotron announcement claimed up to a 20% accuracy improvement and five-times faster inference after post-training.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute“Up to” figures describe a best observed result under the company’s selected comparison and workload, not a universal speedup. The official material available for this release does not provide independent validation sufficient to treat those comparisons as neutral benchmarks. Before choosing a model, measure the exact prompts, context lengths, tools, concurrency, hardware and quality targets used by your application.
Availability and rollout
- October 15, 2026: NVIDIA announced Nemotron 3.5 Super VL and Bolt 74B-A9B; both were stated to be available at announcement.
- October 23, 2026: NVIDIA scheduled Bolt E4B for availability.
- Distribution: Super VL is listed through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM. Bolt 74B-A9B is listed with llama.cpp, Ollama, LM Studio, Unsloth and vLLM support.
The expansion gives developers a clearer split: Super VL for broad multimodal capability, Lightning for fast specialist work and Bolt for local or edge inference. The right choice is determined less by parameter count alone than by modality, memory, latency, privacy and the amount of agent orchestration the application can support.
Quick Recap
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.
