Nemotron 3.5 Super VL: NVIDIA's 120B Open Model Runs Your Agents Locally
NVIDIA shipped Nemotron 3.5 Super VL on October 15, 2026 — a 120-billion-parameter open model with just 12B active parameters, hybrid Mamba-2 and attention MoE architecture, a C-RADIO vision encoder, and up to 1M tokens of context. It takes text, images, and video as input, and it's built for one thing: long-running, specialized AI agents. On OpenRouter it lists for as little as $0.06/$0.35 per million tokens — a fraction of closed-frontier pricing — and you can post-train and self-host it under the OpenMDW-1.1 license.
- 120B total parameters, 12B active — Nemotron 3.5 Super VL is a vision-language mixture-of-experts model with a hybrid Mamba-2/Transformer architecture and a C-RADIO vision encoder. It supports context lengths up to 1M tokens, per its official Hugging Face model card. The OpenRouter listing lists $0.06/$0.35 per 1M input/output tokens as the floor price.
- Built for specialized agents, not general chat — NVIDIA positions the model for "specialized always-on agents" — instruction following, tool use, multistep business workflows, and visual document understanding. The technical blog stresses post-training on your own data via NeMo as the core workflow: dozens of organizations including Scale AI, Perplexity, Atlassian, and CrowdStrike are already post-training it.
- Open weights, datasets, and recipes — Ship includes open weights, open pre-training datasets, and training recipes. License: OpenMDW-1.1, ready for commercial or non-commercial use. Checkpoints: BF16 and NVFP4 (fits on a single DGX Spark).
This article is based on NVIDIA's official blog, its Hugging Face model card, and independent reporting from VentureBeat, all published October 15, 2026. Pricing figures are from OpenRouter as of October 24, 2026. No hands-on testing was performed for this article.
The most interesting AI model launch of October 2026 might not be a closed-frontier one. NVIDIA's Nemotron 3.5 Super VL, released October 15, is a 120-billion-parameter open model with a twist: only 12 billion parameters are active per token, thanks to a hybrid Mamba-2 and attention mixture-of-experts architecture. That design is aimed squarely at one workload — specialized, always-on AI agents that need to run fast and cheap for hours.
Here's what it is, what it can do, and whether it belongs in your stack.
What Nemotron 3.5 Super VL actually is
Per its official Hugging Face model card, the model combines interleaved Mamba-2, mixture-of-experts (MoE), and attention layers with a C-RADIO vision encoder and a vision-language connector. It has 120B total parameters and 12B active parameters, and supports context lengths up to 1M tokens. Inputs are text, images, and video; output is text. It also supports native speculative decoding via multi-token prediction for faster inference, and it has a configurable reasoning mode (off, on, and on-medium) toggled through the chat template.
The OpenMDW-1.1 license means it's ready for commercial or non-commercial use, with weights available in BF16 (full precision) and NVFP4 (quantized to fit on a single DGX Spark). It ships with open weights, open datasets, and training recipes — the full-stack openness NVIDIA has been pushing across the Nemotron line.
Why vision-language matters for agents
NVIDIA's pitch is that specialized agents — claims processing, credit agreement analysis, security operations — need to work across PDFs, charts, tables, images, webpages, and screenshots without routing work between separate text and vision models. VentureBeat's writeup gives the concrete example: an enterprise agent reviewing a credit agreement may need to extract figures from tables, interpret a chart, check a related webpage, and apply the company's own rules before producing an answer.
On performance, NVIDIA claims the model delivers accuracy "comparable to NVIDIA's largest reasoning model at roughly one-fifth the parameter count" and calls it the "fastest open agentic model in its class — 5x faster than the next best" on the Artificial Analysis Intelligence Index. Treat those as vendor claims until independently verified.
The post-training strategy
The most notable part of the launch is not the model but who is already customizing it. Per NVIDIA's announcement, organizations post-training Super VL include Scale AI (for Skydance's vendor credit finance agent and EY's M&A due diligence agent), CrowdStrike (for agentic security operations), Perplexity (for search and agentic tasks), Atlassian, Salesforce, and Mercor (for coding agents). NVIDIA puts it plainly: unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure, and oversee what they build.
Two companion models round out the family: Nemotron 3.5 Bolt 74B-A9B 24GB, built with native 2-bit quantization to fit in 24GB of memory for local agentic workloads on RTX hardware, and the smaller Bolt E4B 4GB for edge and conversational agents.
Pricing: the real story
On OpenRouter, Nemotron 3.5 Super VL lists at $0.06 per million input tokens and $0.35 per million output tokens at the cheapest provider, with mainstream providers like Together, DeepInfra, and CoreWeave in the $0.15/$0.50 range. Compare that to closed-frontier pricing at $2 to $10+ per million input tokens and the appeal is obvious: you can run a specialized agent 10-20x cheaper than a closed model — or self-host the NVFP4 checkpoint on a single workstation if volume justifies it.
Seven providers currently serve the model on OpenRouter: Darkbloom, Crusoe, io.net, Together, DeepInfra, CoreWeave, Baseten, and Venice — a strong sign of ecosystem confidence for a model released just this month.
If you build specialized agents: this is the model family to evaluate for high-volume, document-heavy workloads — post-train on your data, keep it on your infrastructure, and cut per-task costs dramatically versus closed APIs.
If you need frontier reasoning for complex multi-step reasoning or research workloads, keep a frontier closed model in the loop. Specialized agents are this model's job, not everything.
If you self-host and want the best accuracy-per-GB for your AI agents, Nemotron 3.5 Super VL in NVFP4 on a single DGX Spark is now the benchmark to beat.
Bottom line: Nemotron 3.5 Super VL is NVIDIA's clearest statement yet that the next phase of AI value creation is not a bigger general model — it's a specialized, post-trained, self-hosted agent that runs wherever your data lives, at a fraction of closed-model cost. For enterprises looking to build AI agents that read documents, use tools, and respect data-residency requirements, this is the release to watch this quarter.
- NVIDIA Blog — Open From Data Center to Desktop: NVIDIA Expands Nemotron Family (Oct 15, 2026)
- Hugging Face — NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 model card (Oct 15, 2026)
- OpenRouter — Nemotron 3.5 Super VL: API Pricing & Benchmarks (accessed Oct 24, 2026)
- VentureBeat — NVIDIA's new, open Nemotron 3.5 Super VL with post-training beats Claude Opus 5, GPT-6.1 Sol on specialized task benchmarks (Oct 15, 2026)
Prices and benchmark figures are standard published rates as of October 25, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.