NVIDIA’s Nemotron 3.5 Super VL targets visual, tool-using agents

The target is a visual agent that can also use tools

On October 15, NVIDIA released Nemotron 3.5 Super VL, an open-weight model aimed at agents that need to reason across text and visual information. In its release announcement, NVIDIA describes use cases such as reading a document, checking several sources, following a company policy, and calling tools in one task. Its model card says the release has 120 billion total parameters, with 12 billion active per token, and a context window of up to one million tokens.

The model accepts text, images, and video, then returns text. Documents can be provided as page images with accompanying text. That makes it a vision-language model for interpreting material and producing a response, not an image or video generator. NVIDIA says the design combines a Mamba-2 and Transformer mixture-of-experts architecture with a vision encoder and a connector between visual and language representations.

“Super VL” is also distinct from the smaller Nemotron 3.5 Bolt models announced alongside it. NVIDIA describes Bolt 74B-A9B as optimized to fit a 24GB local system and Bolt E4B as a much smaller conversational agent. Those figures do not describe the 120B Super VL checkpoint. The product-family headline spans different hardware tiers; the larger model has its own deployment requirements.

Open weights give developers more control, with a deployment job attached

NVIDIA made base, post-trained, and NVFP4-quantized checkpoints available through repositories and deployment services, including Hugging Face, ModelScope, OpenRouter, and NVIDIA NIM. The model card lists vLLM, SGLang, and TensorRT-LLM, with supported NVIDIA hardware including GB200 and DGX Spark systems plus H100 and H200 GPUs. It prefers Linux. This is a developer release with real inference requirements, not a model that becomes easy to run merely because the weights can be downloaded.

The model card identifies the OpenMDW-1.1 license and says the downloadable model is ready for commercial or noncommercial use. The hosted NVIDIA API trial has separate trial terms. That distinction matters: a team using its own checkpoint is governed by the model license, while a team sending prompts to the hosted endpoint also depends on that service’s terms and operating conditions.

A smaller active-parameter count also does not describe the model file’s storage footprint. Mixture-of-experts routing activates a subset of parameters for each token, but all experts still form part of the stored checkpoint. The NVFP4 version is designed to reduce memory requirements, yet the model card still lists GPU-accelerated platforms. Teams should check the exact checkpoint and runtime before treating “12B active” as a hardware estimate.

The one-million-token context figure also has a boundary: NVIDIA says prompts, visual inputs, conversation history, tool results, and the generated answer share that window. An image or a long tool transcript uses part of the same capacity. A team handling large files should measure how its real inputs fit, rather than count on the headline limit as space reserved only for source documents.

Open weights do not determine where an application sends its requests. A team may run a checkpoint on its own supported GPUs or use a hosted route such as NIM or an API provider. NVIDIA’s model card separates the model license from the hosted API trial terms. If a deployment requires local processing or a particular data region, the team must verify the actual topology and service terms rather than infer those protections from the weight files.

The five-times claim is about a narrow speed comparison

NVIDIA says in its technical blog that Super VL delivers up to five times higher throughput than other open models in its size category on agentic tasks. Its technical blog describes a figure using a 100,000-token input on a Baseten endpoint and reports more than 600 output tokens per second for that chart. NVIDIA also shows a separate test of 10,000 coding tasks on B200 GPUs, where it reports four-times faster task completion. These are different measurements and should not be collapsed into one general promise.

Throughput measures how many outputs a system can produce in a period; task-completion time also depends on how many steps, tool calls, and corrections a workflow needs. A faster token stream does not establish that an agent reaches the right answer more often. NVIDIA’s numbers are useful for understanding the company’s efficiency case, but the post does not show that every customer’s hardware, prompts, or agent harness will reproduce them.

An independent index supports the speed story without proving every use case

Artificial Analysis lists Nemotron 3.5 Super VL at 26 on its Intelligence Index and places it first among 67 models in the comparison class shown on the page when checked on October 23. The evaluator’s current summary also reports a median provider speed of about 428 tokens per second. Those figures offer an outside reference point, but they do not use the same single-GPU Baseten setup in NVIDIA’s five-times claim.

Artificial Analysis describes its measured input support as text and image, while NVIDIA’s model card also lists video input. The index therefore adds an independent view of several reasoning and agent benchmarks, but it does not verify every advertised video capability or reliability in a company’s live workflow. A composite score is not a substitute for testing the exact type of document, video, tools, and failure cases an organization plans to use.

An agent-ready model still needs an agent system around it

NVIDIA’s pitch is for “always-on” and longer-running agents, where a request can read files, call tools, and continue across multiple steps. The model card recommends use-case-specific testing and says deployers are responsible for guardrails and other safety mechanisms. NVIDIA’s blog also describes post-training by partners for specialized workflows, but the reported gains belong to those particular partner tasks and are company-reported examples.

For a team, the first test should separate model behavior from the software wrapper. Give it a read-only task using representative documents, ask it to cite the source page for each extracted value, and compare its result with a human reference. Then test tool calls in a sandbox before allowing writes. These steps will show whether multimodal input helps the chosen workflow and whether the surrounding system keeps the agent’s actions inside the intended boundary.

NVIDIA’s named deployments illustrate why fine-tuning must be kept separate from base-model results. The blog says Scale AI post-trained Super VL for vendor-credit processing and reported a projected cost reduction for that workload; Ramp described a receipt-extraction workflow with its own expected throughput and latency gains. These are examples from partner-specific systems, not measured outcomes for a fresh download or a different company’s process.

The release is most relevant where visual evidence and tools meet

Nemotron 3.5 Super VL is aimed at tasks where a system must interpret text and visual material while using tools: document processing, chart review, video analysis, research, coding, or security triage. Its open weights and quantized checkpoints give infrastructure teams options to adapt or self-host the model, while NIM and listed API providers offer hosted routes.

The release does not make “open” synonymous with effortless, nor “120B” a universal indicator of quality. The useful details are the supported input types, checkpoint, GPU environment, license, API terms, and measured task. NVIDIA has made those details more visible than the slogan. Independent benchmarking provides an additional reference; an organization’s own controlled workflow test still decides whether the model is a good fit.