NVIDIA Releases 120B Open Multimodal Model Built for Always-On Enterprise Agents

Quick Facts

• NVIDIA released Nemotron 3.5 Super VL on Oct. 15, 2026 — a 120B-parameter open multimodal model with 12B parameters active per token and a 1M-token context window.
• Post-trained versions of the model beat Claude Opus 5 and GPT-6.1 Sol on five specialized benchmarks, including medical claims processing and security attack resilience, though NVIDIA has not published full testing protocols for those evaluations.
• Early partners report significant gains: Scale AI cut inference costs 40x and latency 5x for Skydance's vendor credit processing workflow built on the model.

NVIDIA released Nemotron 3.5 Super VL on Oct. 15, 2026, as an open-weights model targeting always-on enterprise agents that need to read documents, analyze images and video, call tools, and write code in a single pass.

The model uses a Mamba-2 mixture-of-experts architecture paired with a C-RADIO vision encoder, a vision projector, and a temporal embedder for video. It carries 512 routed experts per MoE layer, 120 billion total parameters, and 12 billion active per token. Weights are available on Hugging Face in BF16 post-trained, base, and NVFP4 formats.

Speed on Current Hardware

Artificial Analysis measured Super VL at 625 tokens per second on four Blackwell B200 GPUs, according to NVIDIA's developer blog. On a single GPU, NVIDIA claims the model delivers up to 5x higher throughput for agentic tasks compared to open models of similar size.

NVIDIA's LiLiCorr speculative decoding method delivers up to 2.3x faster token generation than decoding without speculation on DGX Spark hardware. The model also uses NVFP4 quantization with quantization-aware distillation, multi-token prediction, and context reuse to reduce inference overhead.

The Benchmark Claims, With Caveats

NVIDIA's post-training results show Super VL outperforming both Claude Opus 5 and GPT-6.1 Sol on five specialized evaluations: document intelligence, security attack resilience, financial receipt extraction, scientific experimentation, and medical insurance claims. Reported gains from post-training range from 9% to 44% across those tasks.

The medical claims result shows the sharpest improvement, with accuracy rising from roughly 28% to 72% after post-training, compared with about 38% for Claude Opus 5 on that evaluation. Security attack resilience climbs from roughly 35% to 76%, compared with about 66% for GPT-6.1 Sol.

The comparisons carry a material qualification. NVIDIA has not supplied task definitions, sample sizes, or testing protocols for these evaluations. The results measure customized systems on specific workflows, not general reasoning ability. Public benchmark trackers still place Claude Opus 5.5 and Gemini 4 Argon ahead of open models on general-purpose reasoning indices as of Oct. 15.

On standard public benchmarks, the picture is mixed. Super VL scores 88.22 on LiveCodeBench v6, ahead of Qwen3.5-122B at 76.7. On MMLU-ProX, Qwen leads 83.2 to 76.91. NVIDIA's own model card notes that Super VL trails GLM-5.3-Flash and DeepSeek V4 Flash on coding and browsing tasks.

Partner Ecosystem

NVIDIA named a broad set of companies already building on Super VL. Scale AI built a vendor credit processing agent for Skydance Corporation, reporting a 40x drop in inference cost and a 5x reduction in latency. EY built a merger-and-acquisition due-diligence agent. AfterQuery adapted the model for Hebbia's credit-agreement analysis work.

Software platforms integrating Super VL include Atlassian for agentic search across Jira and Confluence, CrowdStrike for enterprise security, Dassault Systemes for engineering design and simulation, Perplexity for search, Salesforce for government workflow automation, and Synopsys for semiconductor design verification. Ramp, Dream, Lila Sciences, Mercor, and R1 are also named as early adopters.

What Else Shipped

NVIDIA released two smaller Bolt models alongside Super VL. The Nemotron 3.5 Bolt 74B-A9B 24GB is available now for workstations and local hardware. The Nemotron 3.5 Bolt E4B 4GB, a 4-gigabyte edge build, is scheduled for release Oct. 23, 2026.

For software teams evaluating open-weight alternatives to GPT or Claude for high-volume, repeatable workflows, Super VL's inference economics are the clearest argument. The Skydance result — 40x cost reduction on a defined document-processing task — is a concrete data point. Teams should run their own evaluations before drawing conclusions from NVIDIA's proprietary benchmark charts.