Recommended Free Tools
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
Nemotron 3.5 Super VL is NVIDIA’s open, customizable vision-language model for text, images, documents and video. Its mixture-of-experts design has 120 billion total parameters but activates about 12 billion per forward pass. It can reason over charts, tables, screenshots and long documents, write and inspect code, and call tools through supported agent integrations. The documented NVIDIA NIM interface accepts text, image and video input and returns text; it does not accept audio, and the cited interface does not establish native image or video generation. Running it locally is an infrastructure project rather than a typical consumer-PC installation.
What Nemotron 3.5 Super VL is
Nemotron 3.5 Super VL is a vision-language agent model: it combines a language model with visual encoders and an agent-oriented training set so one system can interpret visual material, reason about it, produce text and interact with external tools. NVIDIA describes it for coding, search, tool-assisted workflows, document question answering, and image and video understanding.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Dell NVIDIA Tesla V100 GPU SXM2 32GB NWWWX by DELL | $854.96 | Buy on Amazon |
| 2 |
|
Gigabyte NVIDIA GeForce RTX 3060 Gaming OC V2 Graphics Card - 12GB GDDR6, 192-bit, PCI-E 4.0,... | $695.00 | Buy on Amazon |
It is a mixture-of-experts (MoE) model. The model contains 120 billion total parameters, while approximately 12 billion are active for an individual forward pass. That distinction matters: total parameters describe the model’s capacity and storage footprint, whereas active parameters affect the computation used for each token. A 120B model still requires serious memory, networking and serving planning even though only a fraction of the experts are selected for each operation.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsThe release is intended to be customizable and is available for commercial or non-commercial use under NVIDIA’s OpenMDW-1.1 license. License permission does not remove the practical obligations of running a large model, such as GPU memory, compatible software and operational monitoring.
#1 Best Overall
- GPU Chipset: NVIDIA
- Memory: HBM2
- Programming Interface: CUDA
- Memory Capacity: 32GB
- Slot Compatibility: SXM2
Which inputs it can actually understand
Text, still images and screenshots
Text and image input are first-class capabilities in the documented NIM interface. A screenshot can be supplied alongside a question, allowing the model to describe visible content, extract text, explain an interface or use the image as evidence while answering a text prompt. As with any vision model, tiny fonts, compression artifacts, occlusion and ambiguous visual context can still produce errors; visual input should be treated as evidence to verify, not an automatic ground truth.
OCR and visual documents
NVIDIA says the training coverage includes optical character recognition (OCR), charts, tables and multipage documents. This enables workflows such as asking a question about a scanned report, extracting fields from a form, or locating a value in one page and relating it to a table on another page.
The important distinction is between reading isolated pages and reasoning across a document. Nemotron 3.5 Super VL is designed to connect information across pages—for example, matching a chart on one page with the underlying table elsewhere—provided the pages are supplied within the request and remain legible. Layout-heavy files, handwriting, unusual typefaces and low-resolution scans should be validated with spot checks.
Charts and tables
Chart understanding is more than OCR. The model must identify axes, legends, units, trends and relationships, then connect those visual elements to a question. NVIDIA reports an 80.8% score on the CharXiv chart-reasoning benchmark (NVIDIA, 2026). That is a benchmark result, not a guarantee that every business dashboard or scientific plot will be interpreted correctly. Ask it to state the units and the visual evidence it used, and compare critical numbers against the source table.
Video
The model includes a temporal video embedder and the NIM interface documents video input. This supports questions about events, sequences and visual changes over time rather than only single frames. Video workloads add practical constraints: clips must be decoded and sampled, long footage consumes context and compute, and a model may miss a brief event if the sampling strategy does not include the relevant frames. For incident review or compliance work, preserve timestamps and verify conclusions against the original recording.
Audio is not an input
NVIDIA’s NIM documentation is explicit: “It accepts text, image, and video input and returns text; it does not accept audio input.” A video file should therefore not be assumed to provide usable speech or sound understanding through this interface. If a workflow depends on dialogue, music or other audio, transcribe or otherwise process the audio with a separate system and pass the resulting text, along with any relevant video frames, to Nemotron.
What it returns
The cited interface returns text. It can describe an image, write code, answer a document question or produce a tool-call decision represented in the agent protocol, but the available documentation does not establish native image or video generation. Do not choose it as an image or video creation model based solely on its ability to understand those modalities.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Agent, reasoning and coding capabilities
Tool use and MCP integration
Nemotron 3.5 Super VL is intended to operate in workflows where a model decides when to use an external capability. NVIDIA documents tool calling and Model Context Protocol (MCP) integration in its NIM onboarding path. In practice, your application supplies tool definitions, accepts the model’s structured request, executes the approved operation and returns the result for another reasoning step. The model does not gain permissions automatically: authentication, allow-lists, input validation, timeouts and audit logs remain application responsibilities.
Search and document question answering
Search-oriented agents can combine retrieved text with screenshots, scans or video frames, then ask the model to reconcile the evidence. For document question answering, a useful pattern is to retain page numbers, table labels and timestamps in the application’s context so an answer can be traced back to its source. Long context helps, but it does not make irrelevant or contradictory material disappear; retrieval and source selection still determine what the model sees.
Reasoning control
Reasoning can be enabled or disabled through the chat template. Disabling it can be appropriate when latency, output length or predictable formatting matters. Enabling it may help with multistep analysis, code interpretation or tool planning, but additional reasoning tokens do not guarantee a correct answer. Test both settings on representative tasks rather than assuming one is universally better.
Software engineering
NVIDIA reports training on software repositories, issue and pull-request pairs, commit diffs, build and continuous-integration logs, and agent trajectories. The model can therefore inspect code, explain a failure, propose a patch, interpret logs or participate in a tool-driven development loop. Repository-specific conventions, dependency versions and security policies still need to be supplied in context. Generated patches should pass the project’s own tests and review process; a coding benchmark score is not proof that an unreviewed change is safe to deploy.
Published evaluation results—and what they do not prove
The following figures are published evaluations attributed to NVIDIA in 2026, unless otherwise noted. They are useful reference points, not service-level guarantees for your data, prompts or tool environment.
Rank #2
- NVIDIA Ampere Streaming Multiprocessors: Building blocks for the world's fastest, most efficient GPUs, the all-new Ampere SM brings twice the FP32 throughput and improved energy efficiency
- 2nd Generation RT Cores - Experience 2x the 1st Generation RT Cores throughput, plus competitive RT and shading for a whole new level of ray-tracing performance
- 【3rd Generation Tensor Cores】Get up to 2X the throughput with structural sparsity and advanced AI algorithms such as DLSS
- Core Clock: 1837MHz
- WINDFORCE 3X Cooler
| Evaluation | Reported result | How to interpret it |
|---|---|---|
| CharXiv chart reasoning | 80.8% | Measures a named chart-reasoning task; performance on a particular dashboard or scientific figure can differ. |
| Python code interpretation | 84% | Indicates capability on the reported code-interpretation evaluation, not guaranteed execution correctness. |
| Hermes Bench | 63.44% | A named agent/tool-use evaluation; actual success depends on your tool schemas, policies and error handling. |
| LangChain behavioral evaluation | 87.05% | A partner or framework evaluation of behavior under its test conditions, not autonomous reliability in every workflow. |
| GPQA-Diamond, BF16 | 89.52% | NVIDIA API Catalog result for the stated precision and benchmark. |
| GPQA-Diamond, NVFP4 | 89.90% | NVIDIA API Catalog result for the stated quantized configuration and benchmark. |
None of these numbers establishes factuality, safe autonomy, resistance to prompt injection, or successful completion of an arbitrary production task. Build an evaluation set from your own documents, charts, videos, code and tool failures. Measure extraction accuracy, citation or page-reference accuracy, latency, cost, refusal behavior and recovery after a tool error.
How the architecture supports multimodal work
The language backbone is a Nemotron-H hybrid combining Mamba-2 and Transformer components in a mixture-of-experts design. A C-RADIO vision encoder turns image information into representations the language system can use. A vision-language connector joins those representations to the language stream, while a temporal video embedder handles information that changes across frames.
The NVFP4 release adds quantization-aware distillation and supports contexts of up to 1 million tokens. It also includes native Multi-Token Prediction speculative decoding, a technique intended to improve generation efficiency by proposing multiple tokens and verifying them. Quantization can reduce memory and improve serving economics, but it may change quality and hardware requirements; compare the exact profile you plan to deploy rather than treating “NVFP4” as a universal speed guarantee.
Context length: what “1 million tokens” means
The API reference lists context lengths up to 1 million tokens, and NVIDIA’s NIM guide specifies a maximum of 1,048,576 tokens for packaged profiles. This is a ceiling for a supported configuration, not a recommendation to place an entire data lake in one prompt.
Images, video frames and document pages consume context or related processing capacity. Very large requests increase latency, memory pressure and the chance that important details receive little attention. A better design usually combines retrieval, page or frame selection, concise metadata and explicit source identifiers. Confirm the maximum context for the particular NIM profile and hardware you deploy; profiles can impose stricter limits than the headline maximum.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Can you run it locally?
Yes, local execution is documented on NVIDIA DGX Spark. That documentation is evidence for the named platform, not evidence that every consumer desktop or laptop can run the model. With 120B total parameters, deployment requires planning for GPU memory, host memory, tensor parallelism, storage, interconnects and serving concurrency.
NIM and serving software
The NIM backend is vLLM 0.31.0. NVIDIA supplies packaged profiles validated for particular GPUs. A profile is therefore part of the deployment specification: selecting a model file without checking its GPU, precision, context and parallelism requirements can lead to an unsupported or unusable setup.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Practical deployment checklist
- Choose the serving target. Decide whether the workload will use an NVIDIA-hosted API, a supported NIM deployment or a documented DGX Spark local installation.
- Match a packaged profile to hardware. Check the profile’s GPU validation, precision, memory requirement, context limit and tensor-parallel configuration before downloading artifacts.
- Plan multimodal preprocessing. Define how PDFs become page images or text, how video is sampled, and how audio is transcribed externally when needed.
- Define the agent boundary. Register only the tools the application needs, validate arguments, enforce authorization and record every call and result.
- Test with production-shaped data. Include poor scans, dense charts, long documents, code repositories, video timing questions and tool failures—not only clean demonstrations.
- Monitor the service. Track queue time, generation latency, GPU memory, context usage, tool-error rates and answer quality after model or profile changes.
Local hosting can offer control over data and customization, but it transfers hardware, upgrades and reliability work to your team. Hosted access reduces infrastructure work while introducing service limits, network dependence and a separate data-governance review.
Performance and cost claims need careful qualification
NVIDIA cites partner-reported improvements for customized workflows: 40-times lower inference cost and five-times lower latency for a vendor-credit workflow, and up to 50% higher throughput with 10-times lower latency for a receipt-processing workflow. These are results for named, customized applications under their reported conditions. They are not general benchmarks for every prompt, batch size, hardware setup or model comparison. Reproduce the measurement with your own traffic before using such figures in a capacity or budget plan.
What it can and cannot do at a glance
| Capability | Supported or established? | Operational qualification |
|---|---|---|
| Text input and text answers | Yes | Documented NIM behavior. |
| Image and screenshot understanding | Yes | Visual quality depends on resolution, layout and prompt. |
| OCR, charts, tables and multipage documents | Yes, as training targets | Verify critical values and page relationships. |
| Video understanding | Yes | Sampling, clip length and temporal detail affect results. |
| Audio input | No in the documented NIM interface | Transcribe or process audio separately. |
| Native image or video generation | Not established by the cited interface | Its documented output is text. |
| Tool calling and MCP | Documented | Your application must execute and secure tools. |
| Local execution | Documented on DGX Spark | Not evidence of broad consumer-hardware support. |
| Million-token context | Up to 1,048,576 tokens in packaged profiles | Check the exact profile and account for multimodal overhead. |
When it is a good fit
Choose Nemotron 3.5 Super VL when one workflow needs several of the following together: visual document extraction, chart or table reasoning, video analysis, code work, long-context synthesis and controlled tool use. Its open and customizable licensing can matter when you need to adapt or host the system rather than rely entirely on a closed endpoint.
It is a weaker fit when the core requirement is audio conversation, native image or video creation, a lightweight laptop deployment, or fully autonomous action without human and software controls. In those cases, use a separate audio or media-generation component, a smaller model, or a deliberately staged system and pass only the resulting text or visual evidence to Nemotron.
Quick Recap
How to evaluate it responsibly
- Separate perception from action. Score OCR, chart reading and video event detection before allowing the model to call tools.
- Require traceability. Preserve page numbers, table names, frame timestamps and tool-result identifiers in the application state.
- Test adversarial inputs. Include prompt injection inside documents, misleading chart legends, malformed files and tools that return errors or partial data.
- Set approval gates. Require human confirmation for financial, destructive, security-sensitive or externally visible actions.
- Compare precision profiles. Measure quality, throughput and memory for the exact BF16, NVFP4 or other profile you will operate.
- Re-evaluate after changes. A new NIM profile, preprocessing pipeline, context limit or tool schema can change behavior even when the model name stays the same.
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.
