What are the strengths of this model?

Open weights under the commercial OpenMDW-1.1 license Unified text, image and video understanding in one model 1M-token context; runs on a single DGX Spark in NVFP4 Early enterprise adoption at Ramp, EY, CrowdStrike and others

What are the weaknesses of this model?

Artificial Analysis Intelligence Index of 26 vs ~58 for closed leaders Most headline benchmarks are NVIDIA's own, some with NeMo Gym scaffolding Best throughput effectively requires NVIDIA GPUs Video input capped at two minutes; no native audio

What are the best use cases?

Document, receipt and PDF agents On-prem compliance and data-residency processing Local high-throughput vision agents Customizable inference on NVIDIA hardware

NVIDIAOpen Source

Nemotron 3.5 Super VL

Name: Nemotron 3.5 Super VL
Author: NVIDIA

Compare this model

NVIDIA's Nemotron 3.5 Super VL is an open 120B-parameter mixture-of-experts vision-language model (12B active) released Oct 15, 2026. It pairs a hybrid Mamba-2 / Transformer design with a C-RADIO vision encoder so one set of weights handles reasoning, coding, tool use, and document and video understanding across text, images and up to two minutes of video. Released under the OpenMDW-1.1 license for commercial use, it runs a 1M-token context, needs 4x H100 or 2x GB200 in BF16, and fits a single DGX Spark in NVFP4.

Share:X Hatena LINE

Parameters

120B (MoE, 12B active)

Context Window

License

OpenMDW-1.1

Release Date

2026-10-15

Benchmark Performance

AA Intelligence Index

26.0

LMArena Elo

—

HLE

—

ARC-AGI-2

—

SWE-bench Verified

71.1

GPQA Diamond

89.5

MMLU-Pro

—

LiveCodeBench

88.2

AIME 2025

—

MATH-500

—

API Pricing

Input Price (per 1M tokens)

$0.15

Output Price (per 1M tokens)

$0.5

Billing Mode: standard

Strengths

•Open weights under the commercial OpenMDW-1.1 license
•Unified text, image and video understanding in one model
•1M-token context; runs on a single DGX Spark in NVFP4
•Early enterprise adoption at Ramp, EY, CrowdStrike and others

Weaknesses

•Artificial Analysis Intelligence Index of 26 vs ~58 for closed leaders
•Most headline benchmarks are NVIDIA's own, some with NeMo Gym scaffolding
•Best throughput effectively requires NVIDIA GPUs
•Video input capped at two minutes; no native audio

Use Cases

•Document, receipt and PDF agents
•On-prem compliance and data-residency processing
•Local high-throughput vision agents
•Customizable inference on NVIDIA hardware

Deep Analysis

Parameters

120B (12B active MoE)

Mamba-2 + Transformer hybrid; 512 routed experts/layer; C-RADIO vision encoder

Context Window

1M tokens

Text, images (JPEG/PNG) and video (MP4 up to 2 min); text output

License

OpenMDW-1.1

Open weights on Hugging Face, ModelScope, OpenRouter; commercial + non-commercial

GPQA-Diamond

89.52%

NVIDIA-reported; strong for an open VL model

LiveCodeBench v6

88.22%

NVIDIA-reported agentic coding

SWE-bench Verified

71.10%

NVIDIA-reported; trails top closed coders

AA Intelligence Index

Third-party (Artificial Analysis); well behind ~58 closed leaders

Release

Oct 15, 2026

With Bolt 74B-A9B (24GB local) and Bolt E4B (4GB)

Strengths

・Open-weight and commercially licensed (OpenMDW-1.1), so you can self-host, fine-tune and ship it without the usual open-model restrictions - and NVIDIA published the training data and recipes too.
・Unified multimodal agent model: one set of weights handles reasoning, coding, tool use, document and video understanding, instead of chaining a text model to a separate vision model.
・Runs very long context cheaply: a hybrid Mamba-2/Transformer design holds a 1M-token window at low memory, and NVFP4 lets it run on a single DGX Spark.
・Real enterprise traction at launch: Ramp (receipt parsing, ~10x latency improvement), EY, CrowdStrike, Perplexity and Salesforce reported production use.

Weaknesses

・Benchmark reality check: the Artificial Analysis Intelligence Index is 26 versus ~58 for the leading closed models, so it is an efficient specialist, not a frontier replacement.
・Most headline numbers are NVIDIA's own, some produced with NeMo Gym scaffolding, so they are not directly comparable to other vendors' results.
・Heavy on NVIDIA hardware: the BF16 build needs 4x H100 or 2x GB200; getting the best speed effectively means buying NVIDIA GPUs.
・Multimodal input is capped at 2-minute video and text output only - no native audio or speech.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GPT-5.4 (multimodal)	N/A	Higher (SWE-bench)	N/A	Closed
Qwen3-VL / Kimi K3	N/A	N/A	N/A	Open
Nemotron 3.5 Lightning	N/A	N/A	N/A	Open (text-only)

Overview

Nemotron 3.5 Super VL is NVIDIA's open-weight vision-language model for agents, released on October 15, 2026. It is a 120-billion-parameter mixture-of-experts model that activates 12 billion parameters per token, pairing a hybrid Mamba-2 / Transformer architecture with a C-RADIO vision encoder. The pitch is a single model that reads text, images, documents and short video and then reasons, codes, calls tools and does document and video understanding - rather than bolting a vision model onto a text model. NVIDIA released it under the OpenMDW-1.1 license, which permits commercial and non-commercial use, and published the training data and recipes alongside the weights. Reported scores include GPQA-Diamond 89.52%, LiveCodeBench v6 88.22%, BrowseComp 78.0% and SWE-bench Verified 71.10%. The honest caveat is the third-party picture: on the Artificial Analysis Intelligence Index it scores 26, well behind the ~58 of leading closed models. Think of it as a fast, customizable model for NVIDIA hardware, not a frontier-model replacement. It launched with two smaller siblings - Bolt 74B-A9B (2-bit, fits 24GB) and Bolt E4B (4GB) for voice and embedded use.

Benchmarks & Performance

Detailed Comparison

Community Feedback

Reaction split along predictable lines. Builders on NVIDIA stacks liked that one model replaced a text-plus-vision chain and that the weights, recipes and training data were all published. The skeptical read, led by The Decoder, was that the 26-point Artificial Analysis score is far below the top closed model and that the open-model push is self-serving because the best speed needs NVIDIA hardware. * **What people liked.** The single-model unification (reasoning + vision + tools) and the published datasets/recipes for fine-tuning. EXO Labs found that on DGX Spark no other tested model was both more accurate and faster, which is a real local-inference win. * **What people pushed back on.** The gap to closed leaders on the AA index (26 vs ~58), and that some NVIDIA benchmarks used NeMo Gym scaffolding. The throughput claim (up to 5x) is single-GPU and NVIDIA-specific. * **Practical adoption.** Early production reports came from enterprises: Ramp (receipt parsing, ~10x latency improvement, up to 50% higher throughput), Scale AI (vendor credit processing, 40x cost reduction, 5x latency drop) and AI21 (threat-intel extraction, 40% accuracy gain).

Use Cases

### Where Nemotron 3.5 Super VL earns its keep **1. Document, receipt and PDF agents** * **Example:** A finance bot that parses unstructured receipt images into a structured JSON schema. * **Why it works:** Ramp reported ~10x latency improvement and up to 50% higher throughput versus prior setups, and the 1M context swallows long documents whole. **2. On-prem compliance and data residency** * **Example:** A bank processes statements and contracts on its own GPUs for regulatory reasons. * **Why it works:** OpenMDW-1.1 allows commercial use and self-hosting, so sensitive documents never leave the building, and the published recipes let the team fine-tune. **3. Local high-throughput vision agents** * **Example:** A DGX Spark box running receipt or chart understanding at the edge. * **Why it works:** NVFP4 fits on a single DGX Spark, and EXO Labs found it was both more accurate and faster than other tested models on that hardware.

Latest News

**Release (October 15, 2026).** NVIDIA released Nemotron 3.5 Super VL, an open-weight 120B/12B MoE vision-language model for complex agentic workflows, under the OpenMDW-1.1 license. It combines interleaved Mamba-2 state-space layers, sparse MoE routing and a C-RADIO vision encoder, accepting text, images and up to two minutes of video with configurable reasoning (off / medium / on). **Benchmarks (NVIDIA-reported).** GPQA-Diamond 89.52%, LiveCodeBench v6 88.22%, BrowseComp 78.0%, SWE-bench Verified 71.10%, SWE-bench Pro 51.78%, GraphWalks 89.71%, MathVision 78.16%, MMLongBench-Doc 62.82%, IFBench 70.86%, Video-MME v2 32.85%, CharXiv 80.8%. Third-party, the Artificial Analysis Intelligence Index is 26 versus ~58 for leading closed models. **Availability and siblings.** Weights, cookbooks and eval recipes are on Hugging Face, NVIDIA NIM and GitHub (NeMo Gym). It also shipped Bolt 74B-A9B (2-bit, fits 24GB for local agents) and Bolt E4B (4GB, for voice and embedded use). Hosted API pricing was about $0.15 / $0.50 per 1M tokens across six providers as of Oct 21, 2026. **Enterprise adoption.** Ramp (receipt parsing), Scale AI (vendor credit processing, 40x cost reduction), EY, CrowdStrike, Perplexity and Salesforce reported production use.

NVIDIA released it under the OpenMDW-1.1 license, which permits commercial and non-commercial use, and published the training data and recipes alongside the weights. Reported scores include GPQA-Diamond 89.52%, LiveCodeBench v6 88.22%, BrowseComp 78.0% and SWE-bench Verified 71.10%. The honest caveat is the third-party picture: on the Artificial Analysis Intelligence Index it scores 26, well behind the ~58 of leading closed models. Think of it as a fast, customizable model for NVIDIA hardware, not a frontier-model replacement. It launched with two smaller siblings - Bolt 74B-A9B (2-bit, fits 24GB) and Bolt E4B (4GB) for voice and embedded use.

Sources

Analysis generated: 2026-10-22