Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesiTechGuides 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 announced Nemotron 3.5 Super VL on October 15, 2026, alongside two smaller Nemotron 3.5 Bolt models aimed at local and edge hardware. Super VL is a 120-billion-parameter mixture-of-experts model with 12 billion active parameters per token; Bolt 74B-A9B is designed for systems with 24 GB of memory, while Bolt E4B targets 4 GB edge devices and is scheduled for October 23, 2026.
What NVIDIA released
The release creates two branches of the Nemotron 3.5 family. Super VL is the high-capability model for multimodal, tool-using agents. Bolt is the efficiency branch, intended to bring useful agent and conversational workloads to machines with much less memory.
| Model | Positioning | Memory target | Availability announced October 15, 2026 |
|---|---|---|---|
| Nemotron 3.5 Super VL | 120B-total, 12B-active multimodal mixture of experts | Data-center and workstation-class deployments | Available immediately |
| Nemotron 3.5 Bolt 74B-A9B | Local efficiency model | 24 GB systems | Available immediately |
| Nemotron 3.5 Bolt E4B | Compact conversational edge model | 4 GB edge hardware | Scheduled for October 23, 2026 |
The “open” designation for Super VL refers to NVIDIA’s released model and weights, but it should not be assumed to mean that every component has an identical open-source license. The API reference describes the model as ready for commercial or non-commercial use; organizations should still read the license accompanying the checkpoint they download.
Recommended Free Tools
What Nemotron 3.5 Super VL is
Super VL combines reasoning, instruction following, coding, tool use and multimodal understanding in one model. The NVFP4 release accepts text, images and video and generates text, so an agent can inspect a document or video, reason over it, call a tool and return a structured answer without switching between separate language and vision models.
#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
Its 120B parameter count is the total size of the mixture-of-experts system. Only about 12B parameters are activated for each token. That sparse design is intended to provide much of a large model’s capability while keeping per-token computation below that of a dense 120B model.
The technical architecture interleaves Mamba-2 layers, mixture-of-experts layers and attention layers. A C-RADIO vision encoder and a vision-language connector handle visual input. This combination is aimed at long, multistep workflows rather than only short chat responses.
Capabilities that matter for agents
Long context and multimodal input
The NVFP4 checkpoint supports up to a 1-million-token context window. That capacity is useful for large document sets, extended customer histories, codebases or long videos, although actual usable context depends on the serving stack, memory and workload. Inputs can include text, images and video; outputs are text.
Rank #2
- 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
Native speculative decoding
Super VL includes native multi-token prediction for speculative decoding. A compatible serving system can use those predicted token blocks to reduce decoding time, potentially improving responsiveness without changing the model’s answers. The exact gain depends on hardware, batching and acceptance rates.
Tool use and structured workflows
NVIDIA positions the model for customer service, research, document processing, coding and tool-using agents. These are workflows in which the model must retain state, interpret unstructured material, decide on an action and produce a result in several steps.
Reported workflow and benchmark results
NVIDIA’s developer coverage reports more than 600 output tokens per second on an Artificial Analysis agentic-task result using a Baseten endpoint. It also reports 63.44% on Hermes Bench and 87.05% on LangChain evaluations, comparing Nemotron 3.5 Super VL with Nemotron 3 Ultra.
Rank #3
- Powered by the NVIDIA Blackwell architecture and DLSS 4. System Requirements: Minimum 850W PSU with 16-pin 12V-2x6 (12VHPWR) connector required. Verify before purchasing.
- Military-grade components deliver rock-solid power and longer lifespan for ultimate durability. Compatibility: 348mm (13.7") length, 3.6 slots, 4.3 lbs. Confirm case clearance and slot spacing. GPU bracket included.
- Protective PCB coating helps protect against short circuits caused by moisture, dust, or debris
- 3.6-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
Those figures are vendor-reported results, not independent testing. Endpoint configuration, quantization, batching, prompt length and measurement methodology can materially change throughput and accuracy, so they should be treated as reference results rather than a guarantee for a local installation.
NVIDIA also cites partner-reported workflow outcomes. Scale AI post-trained the model for vendor-credit processing at Skydance and reported 40-times lower inference cost and five-times lower latency for that workload. Ramp expects up to 50% higher throughput and 10-times lower latency when converting receipt images into structured JSON. These results apply to the partners’ specific systems and post-training, not automatically to an unmodified checkpoint.
How Bolt differs from Super VL
Bolt addresses the practical problem that a 120B-class multimodal model is not a realistic fit for many local computers. The 74B-A9B edition is aimed at 24 GB systems, while E4B is designed for 4 GB edge hardware and conversational use.
Rank #4
- 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
| Decision factor | Super VL | Bolt 74B-A9B | Bolt E4B |
|---|---|---|---|
| Primary role | High-capability multimodal agents | Local deployment with a larger edge-oriented model | Compact conversational workloads |
| Parameter information published in the release | 120B total, 12B active | 74B total, 9B active | The release identifies the E4B edition; a total/active split is not stated |
| Target memory | Supported data-center and workstation configurations | 24 GB | 4 GB |
| Release timing | Immediate | Immediate | October 23, 2026 |
Choose Super VL when visual understanding, long context, complex reasoning or tool orchestration is the priority and you can provide the required accelerator. Choose Bolt 74B-A9B when local execution on a 24 GB machine matters more than maximum capability. Choose Bolt E4B when the device has only about 4 GB available and the workload is primarily conversational.
Hardware needed for Super VL
NVIDIA’s API reference lists these deployment configurations for the quantized NVFP4 checkpoint:
| Configuration | Deployment description |
|---|---|
| One DGX Spark | Supported single-system configuration |
| One H200 | Supported accelerator configuration |
| Two H100 GPUs | Supported multi-GPU configuration |
| One GB200 | Supported high-end configuration |
These entries describe configurations NVIDIA lists as supported; they are not a universal minimum for every quantization or serving setup. Memory for weights is only one requirement. The runtime also needs space for activations, the key-value cache, multimodal inputs and the chosen context length. A one-million-token context can require substantially more memory than a short prompt.
Best Value
- 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
Ways to run the model
NVIDIA distributes the model through its documentation and a Hugging Face repository. The official materials provide examples for Transformers, vLLM, SGLang and Docker, including an OpenAI-compatible local server path.
- Select the checkpoint and precision. For the published hardware guidance, use the NVFP4 quantized Super VL checkpoint rather than assuming an unquantized version will fit.
- Choose a serving framework. Transformers is suited to direct Python integration; vLLM and SGLang provide production-oriented serving options; Docker packages the runtime and dependencies.
- Check accelerator memory and software support. Confirm that the GPU, driver, CUDA stack and framework version support the selected quantization and multimodal operations.
- Start with a modest context length. Increase it only after measuring memory use, latency and cache growth on representative prompts.
- Expose an API only after validation. If using the OpenAI-compatible server path, test text, image and video requests separately and verify tool-call schemas before connecting an autonomous agent.
Training data and language coverage
NVIDIA says the released Indic-language dataset contains 2.8 trillion tokens across 22 Indian languages and was developed with IIT Madras and AI4Bharat. That data contribution is relevant for teams evaluating multilingual support, but it does not by itself establish equal quality across all 22 languages; language-specific testing remains necessary.
Commercial use and practical caveats
The API reference states that Super VL is ready for commercial or non-commercial use. Before deployment, check the exact license bundled with the checkpoint, the obligations for redistribution and any separate terms for hosted services or third-party components.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →- Reported speed numbers come from specified endpoints and may not match a local server.
- Long context increases memory consumption and can reduce throughput.
- Video and image processing add preprocessing cost and may change latency substantially compared with text-only prompts.
- Tool-using agents need application-level safeguards, including permission boundaries, input validation and logging; model support for tool calls does not make an agent automatically safe.
Bottom line
Nemotron 3.5 Super VL is NVIDIA’s 120B-total, 12B-active multimodal model for demanding always-on agents, with one-million-token context support and native speculative decoding. The Nemotron 3.5 Bolt 74B-A9B and Bolt E4B editions trade capability for much lower memory requirements, bringing the family to 24 GB local systems and 4 GB edge hardware. The right choice is therefore determined first by available memory and workload complexity, then by the latency, context and multimodal features the application actually needs.
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.
