Nemotron 3.5 Super VL vs Nemotron 3 Super: The Multimodal Upgrade

Switch to Nemotron 3.5 Super VL if your pipelines require image, document-page, or video processing, or if you need the OpenMDW-1.1 license. Stay on Nemotron 3 Super if your stack is text-only and already tuned for your tasks.

At Layer3Labs, we build and run automated workflows inside client environments, where changing model checkpoints requires testing against production prompts. Both models feature an identical parameter footprint of 120 billion total parameters and 12 billion active parameters per token.

NVIDIA released Nemotron 3 Super on March 11, 2026 as a text-only model. On October 15, 2026, Kari Briski announced Nemotron 3.5 Super VL in an NVIDIA blog post, adding image and video input at the same 120B size.

On OpenRouter, list pricing as of October 2026 starts at $0.10 per million input tokens and $0.40 per million output tokens for Nemotron 3.5 Super VL, compared to $0.08 input and $0.45 output for Nemotron 3 Super. NVIDIA's published scores for the two models use different benchmark versions and harnesses, so they are not a clean head-to-head.

Nemotron 3.5 Super VL vs. Nemotron 3 Super: Side-by-Side

Are you one of these vendors? Update your listing

Which One to Use

Switch to Nemotron 3.5 Super VL if your workload requires multimodal inputs such as images, scanned document pages, or video clips, or if your organization requires the OpenMDW-1.1 license.

Stay on Nemotron 3 Super if your software handles text exclusively and is already tuned on your current prompt templates.

Both models feature 120 billion total parameters with 12 billion active parameters per token. NVIDIA's published scores for the two use different benchmark versions and harnesses, so they are not a clean head-to-head.

NVIDIA publishes no VRAM figure for Nemotron 3.5 Super VL, and its listed deployments start at 1x DGX Spark, 1x H200 or 2x H100. If you need a model packaged to fit 24GB of memory, evaluate Nemotron 3.5 Bolt instead.

Our recommendation would flip if NVIDIA updates Nemotron 3 Super with vision connectors or if your own tests show Nemotron 3.5 Super VL doing worse than Nemotron 3 Super on your text tasks.

What Changed Between Nemotron 3 Super and 3.5 Super VL

The main change from Nemotron 3 Super to Nemotron 3.5 Super VL is image and video input.

Nemotron 3 Super handles only text inputs and produces text outputs. In contrast, Nemotron 3.5 Super VL accepts text, JPEG and PNG images, and MP4 video clips up to two minutes long. Documents go in as page images with text, and NVIDIA's blog says the model handles OCR, tables, charts, web pages and PDFs.

The licensing terms also shifted between generations. Nemotron 3 Super ships under the NVIDIA Nemotron Open Model License. Nemotron 3.5 Super VL adopts the OpenMDW-1.1 license. NVIDIA also released open datasets and training recipes with the model.

Architecturally, both systems employ a Mamba-2 and Transformer hybrid Mixture of Experts (MoE) design with Multi-Token Prediction (MTP). Nemotron 3.5 Super VL adds a C-RADIO vision encoder and a vision-language connector. Its card lists 512 routed experts per MoE layer. A shared MTP head provides native speculative decoding during inference.

NVIDIA trained Nemotron 3.5 Super VL on approximately 26 trillion tokens across pre-training and continued pre-training, including 100 billion tokens of Indic data and 67 billion tokens of software-engineering data. This foundation was followed by roughly 3.8 million post-training samples and reinforcement learning (RL) across dozens of environments.

NVIDIA Benchmark Metrics and Test Harness Variations

Published benchmarks reported by NVIDIA show higher scores for Nemotron 3.5 Super VL, but differing evaluation harnesses prevent a direct head-to-head comparison.

The benchmark figures for both models come from NVIDIA's own model cards. On Graduate-Level Google-Proof Q&A (GPQA), NVIDIA published an 89.52 BF16 score and an 89.90 NVFP4 score on GPQA-Diamond for Nemotron 3.5 Super VL. In contrast, Nemotron 3 Super has a published GPQA score of 79.23 tested under a no-tools configuration.

Software engineering benchmarks show a similar variation in testing methodology. Nemotron 3.5 Super VL achieved a vendor-reported 71.10 in BF16 and 71.07 in NVFP4 on SWE-bench Verified. Nemotron 3 Super scored 60.47 on SWE-bench Verified using the OpenHands evaluation harness. On command-line tasks, Nemotron 3.5 Super VL scored 60.67 in BF16 on Terminal-Bench 2.1, whereas Nemotron 3 Super scored 31.00 on Terminal Bench Core 2.0.

Nemotron 3.5 Super VL reached 31.79 on text-only Humanity's Last Exam (HLE) and 29.20 across the full set, compared to 18.26 for Nemotron 3 Super without tools. On BrowseComp, Nemotron 3.5 Super VL recorded 78.00 in BF16, while Nemotron 3 Super reached 31.28 on BrowseComp with Search. Third-party testing reported by Let's Data Science notes an Artificial Analysis Intelligence Index score of 26.5 for Super VL.

Where to Run Each Model and What It Costs

Hosted application programming interface (API) rates for Nemotron 3.5 Super VL begin at $0.10 per million input tokens on OpenRouter.

As of October 2026, OpenRouter lists Nemotron 3.5 Super VL with an input price of $0.10 per million tokens, an output price of $0.40 per million tokens, and cached input at $0.08 per million tokens. OpenRouter also lists a free tier under the model ID nvidia/nemotron-3.5-super-vl:free. Individual host pricing varies, with Crusoe and DeepInfra charging $0.10 input and $0.40 output, Together and CoreWeave charging $0.15 input and $0.50 output, and Venice charging $0.25 input and $1.25 output. On OpenRouter, Nemotron 3 Super lists at $0.08 per million input tokens and $0.45 per million output tokens.

Context window support differs by hosting provider. Both models support context windows up to 1 million tokens. However, the build.nvidia.com API and most OpenRouter providers serve requests at a 262,144-token limit, while CoreWeave serves the full 1,048,576 tokens.

For Nemotron 3 Super BF16, NVIDIA specifies a minimum configuration of 8x H100-80GB GPUs, or 2 GPUs on B200 or B300 hardware. NVIDIA has not published a specific video random-access memory (VRAM) number for Nemotron 3.5 Super VL. Deployment documentation lists support across 1x DGX Spark, 1x H200, 2x H100, or 1x GB200 systems, as well as local setups on RTX Spark and RTX Pro 6000 workstations. Our guide to running Nemotron locally covers those local setups.

→ Compare per-token pricing for every AI model

Migration Considerations for Production Workflows

Moving an agent from Nemotron 3 Super to Nemotron 3.5 Super VL means re-testing your prompts and reasoning settings before switching traffic.

Nemotron 3.5 Super VL supports configurable reasoning via the enable_thinking parameter in its chat template, alongside support for function calling and structured outputs. Engineering teams must test whether existing system prompts intended for Nemotron 3 Super cause unwanted verbosity or parsing errors when thinking features are active.

Ingestion pipelines must be adapted when adding visual capabilities. Rather than extracting raw text from PDFs and passing unformatted strings, pipelines send each document page as a JPEG or PNG image along with its text.

Teams must also verify token limits across their hosting stack before switching traffic. Because default configurations often cap context at 262,144 tokens, long-running agent loops that rely on extended history must be tested against target endpoint limits. For self-hosting, NVIDIA's card lists vLLM, SGLang and TensorRT-LLM among the runtimes for the NVFP4 checkpoint. Have whoever signs off on model licenses review OpenMDW-1.1, because Nemotron 3 Super shipped under the NVIDIA Nemotron Open Model License and Nemotron 3 teams may have approved only that license.

Where Nemotron 3.5 Bolt Fits in the Model Stack

Teams that require local execution on consumer or workstation hardware should evaluate Nemotron 3.5 Bolt instead of either 120B model.

Announced alongside Nemotron 3.5 Super VL on October 15, 2026, Nemotron 3.5 Bolt is a 74-billion total parameter model with 9 billion active parameters. NVIDIA created Bolt by compressing Nemotron 3.5 Super VL BF16 using Star Elastic, reducing the routed expert top-K count from 22 to 9.

Bolt fits inside a 24GB memory footprint through native 2-bit quantization, applying IQ2-XXS at 2.0625 bits per element to its experts while retaining NVFP4 for Mamba layers. It supports the same input modalities as Super VL, including text, images, and video clips up to two minutes, with a context window of up to 256K tokens under the OpenMDW-1.1 license. Speculative decoding runs via DFlash with the LiLiCorr drafter for 7 speculative tokens.

Deployment for Bolt targets local hardware such as GeForce RTX 5090 GPUs, RTX Spark laptops, DGX Spark, and RTX PRO workstations across llama.cpp, Ollama, LM Studio, Unsloth, and vLLM. NVIDIA claimed Bolt is the top open model in its class on PinchBench, although its reported score of 87.82 sits behind Qwen3.6-35B-A3B BF16 at 88.07 on the model card table. Bolt is not available on OpenRouter as of October 2026, while a smaller Bolt E4B 4GB model is scheduled for release on October 23.

The Verdict

Select Nemotron 3.5 Super VL if your production agents need to ingest images, scanned documents, charts, or short video clips, or if your organization requires the OpenMDW-1.1 license (NVIDIA also released open datasets and training recipes with this model). Its 120B-A12B architecture maintains parity with the parameter footprint of Nemotron 3 Super while introducing vision encoding.

Remain on Nemotron 3 Super if your application processes plain text exclusively and is already tuned on your current prompt templates. Both models activate 12 billion parameters per token, and OpenRouter list prices as of October 2026 are close: from $0.10/$0.40 per 1M input/output tokens for Super VL versus $0.08/$0.45 for 3 Super.

Deploy a side-by-side evaluation on your proprietary test dataset to measure latency and output quality before committing production traffic to the new model.