Nemotron 3.5 Super VL

NVIDIA Nemotron 3.5 Super VL is a 120B-parameter MoE model with 12B active parameters, supporting up to a 1 million token context, aimed at document understanding, computer use and agent scenarios.

In a developer blog post, NVIDIA introduced a vision-language model called Nemotron 3.5 Super VL, aimed at multimodal agent scenarios.

The default models of multiple Providers have changed: Cortecs, CrossModel, Eden AI, Google Gemini, GitHub Copilot, Google Vertex AI, LLM Gateway, Ofox, OpenCode Go, and Prime Inference now default to Gemini 4 Argon; Fireworks AI, Nebius Token Factory, and Requesty switch to Muse Spark Open; CoreWeave, DigitalOcean, NanoGPT, and Nvidia adopt Nemotron 3.5 Super VL; Aixy and Neon use Claude Haiku 5.5; Agnes AI, Ambient, IteraCompute, Poe, Scaleway, and SenseNova also have their own adjustments.

Nemotron 3.5 Super VL is priced on OpenRouter at $0.10 per million input tokens and $0.40 per million output tokens, offered by five hosts, four of which only serve the 262,144-token window.

On the Gauntlet, Nemotron 3.5 Super VL fully passes 41 of 99 checks, with its score capped at 0.599 because its 3D panel rendered as an empty canvas.

During the Forge 2.0 Build 1 run, Nemotron 3.5 Super VL's prompt token count dropped abruptly from 146,368 to 7,394 and never exceeded 67,997 afterward; goose did not discard history, but OpenRouter's `middle-out` truncation strategy may be related.

In the tweet, NVIDIA emphasized that the model targets "a range of agent workflows."

Nemotron 3.5 Super VL is distributed through Ollama, enabling individual developers and small teams to trial this vision-language model at low cost locally or in private environments.

The 5x throughput improvement is an upper-bound value measured by the official team under specific benchmarks and configurations, and actual performance will vary by task type, batch size and hardware environment.

Nvidia's official data claims that Nemotron 3.5 Super VL can achieve up to 5 times higher agent throughput on a single GPU than open-source models of the same scale.

The 5x throughput improvement is currently only an official claim, lacking third-party benchmark reproduction.

Nemotron 3.5 Super VL supports four types of input—text, images, documents, and video—and has built-in reasoning and tool-calling capabilities.

Officials did not disclose the specific benchmark scores for each modality of Nemotron 3.5 Super VL, the specific VRAM requirements, or the license details.

Currently disclosed official information centers on availability; parameters such as the specific pricing, call rate limits and context window of Nemotron 3.5 Super VL on Perplexity still require consulting the official documentation.

Nemotron 3.5 Super VL's backbone was initialized from Nemotron 3 Super Base, distilled 1T tokens via Nemotron 3 Ultra Base, extended to a 1M context, then underwent 67B tokens of software engineering mid-training, and only then was vision training done.

Nemotron 3.5 Super VL uses a Mamba-2/Transformer hybrid MoE architecture, with 512 routed experts per MoE layer, paired on the vision side with a C-RADIO encoder, supporting text, JPEG/PNG images, and MP4 video up to 2 minutes long, with a context limit of 1M tokens and native support for multi-token prediction.

The Nemotron 3.5 Super VL repository is 75.1 GB in size, spread across 88 safetensors shards, and a single consumer-grade card cannot accommodate a 120B model.

The RTX 5090 has only 32 GB of VRAM, which is not enough for Nemotron 3.5 Super VL; even the UD-IQ1_M quantized version requires 46.1 GB.

Nemotron 3.5 Super VL uses the OpenMDW 1.1 license, which grants unrestricted rights to dispose of the model materials, has no field-of-use clauses, and does not restrict outputs; redistribution requires retaining the notice, and you cannot sue claiming the model infringes patents or copyright.

Nemotron 3.5 Super VL replaces a pure attention decoder with Mamba state-space blocks, uses a self-developed RADIO vision encoder, and natively supports NVFP4 four-bit quantization on the Blackwell architecture.