What Is Nemotron 3.5 Super VL? 120B Open Agent Model

What Is Nemotron 3.5 Super VL? A 120B Open Vision Agent Running on 12B Active Parameters

Nemotron 3.5 Super VL is an open-weight multimodal mixture-of-experts model from NVIDIA. It combines text, image, document and video understanding with reasoning and tool use in one model built for long-running AI agents. The model has 120B total parameters, activates 12B parameters per token and supports a context window of up to 1M tokens. NVIDIA released it on October 15, 2026, with BF16 and NVFP4 checkpoints. It scores 26 on the Artificial Analysis Intelligence Index, which ranks it #1 of 67 models in its comparison class.

Key Takeaways

• Size: 120B total parameters and 12B active per token, in a hybrid Mamba-Transformer MoE design.
• Inputs: text, images, documents, GUIs and video. Output is text only. The context window reaches up to 1M tokens.
• Benchmarks: it scores 26 on the Artificial Analysis Intelligence Index, beating the 550B Nemotron 3 Ultra (23) and doubling the earlier Nemotron 3 Super (13).
• Speed: NVIDIA claims up to 5x higher throughput than open models in its size class. Artificial Analysis measures 196 output tokens per second.
• License and weights: the weights are open under OpenMDW-1.1, according to Artificial Analysis, and published in BF16 and NVFP4.
• Local use: NVIDIA publishes an official recipe for running it on one 128GB DGX Spark.

What Is Nemotron 3.5 Super VL?

Nemotron 3.5 Super VL is the vision-language member of NVIDIA's Super tier in the Nemotron family. NVIDIA released a text-only 120B Nemotron 3 Super in March 2026. Super VL keeps the same base size and adds a vision encoder, so one model can read a PDF, interpret a chart, look at a screenshot and call tools without passing context to a separate perception model.

NVIDIA's launch announcement says the model was built with the Nemotron Coalition. NVIDIA describes its accuracy on multimodal agentic tasks as "comparable with NVIDIA's largest reasoning model at roughly one-fifth the parameter count." The model is fully released, not a preview. Two edge models launched alongside it: Nemotron 3.5 Bolt 74B-A9B 24GB, available now, and Bolt E4B 4GB, scheduled for October 23.

NVIDIA divides work across the family by role. Nemotron 3 Ultra handles the hardest text-only planning. Nemotron 3.5 Lightning, released in August, handles high-volume text steps. Super VL covers the multistep work that needs visual understanding and tool use together.

Anchor definition: Nemotron 3.5 Super VL is NVIDIA's 120B-parameter, 12B-active open multimodal model that merges vision, reasoning and tool calling into one agent backbone with a 1M-token context.

Nemotron 3.5 Super VL at a Glance

Attribute	Detail
Developer	NVIDIA (with the Nemotron Coalition)
Type	Open-weight multimodal reasoning model (vision-language)
Release date	October 15, 2026
Parameters	120B total, 12B active per token
Architecture	Hybrid Mamba-Transformer mixture-of-experts with Multi-Token Prediction
Vision encoder	C-RADIOv4
Input modality	Text, images, documents, GUIs, video (Artificial Analysis lists text and image)
Output modality	Text, with JSON output and tool calling
Context window	Up to 1M tokens
Official checkpoints	BF16, NVFP4 (NVFP4 checkpoint about 75.1GB)
License	OpenMDW-1.1 (per Artificial Analysis)
Pricing	Free weights. Hosted pricing tracked at about $0.20 input / $0.575 output per 1M tokens. NVIDIA list price not yet confirmed
Recommended sampling	temperature 1.0, top_p 0.95
Availability	Hugging Face weights, NVIDIA NIM, NeMo post-training, day-0 support in vLLM and llama.cpp

How Nemotron 3.5 Super VL Works / What Makes It Different

The design goal is throughput for agents that run for hours or days. Several named techniques support that goal:

• Hybrid Mamba-Transformer MoE. Mamba state-space layers sit alongside attention layers, and only 12B of the 120B parameters fire per token. According to NVIDIA's technical blog, this gives up to 5x higher throughput on a single GPU than other open models in its size category.
• Multi-Token Prediction (MTP). A built-in MTP head drafts several candidate tokens per forward pass.
• NVFP4 KV Cache. 4-bit weights and activations run on Blackwell GPUs. NVFP4 quantization of the KV cache also shrinks its HBM footprint, which leaves room for more concurrent long-context sessions and reduces cache eviction in multi-turn agent loops.
• LiLiCorr Speculative Decoding. This speculative decoding module speeds up generation by up to 2.3x on DGX Spark.
• Context Caching. Attention caches and Mamba states are reused across repeated agent calls.

The architecture also produced a sibling model. According to a community breakdown of NVIDIA's technical overview, Nemotron 3.5 Bolt 74B-A9B is not trained from scratch. It is Super VL compressed with NVIDIA's Star Elastic framework (120B to 74B total, 12B to 9B active), then distilled and quantized to fit in 24GB.

The measured speed matches NVIDIA's framing. Artificial Analysis recorded about 630 tokens per second at 100K-token input, roughly 5x the speed of Inkling Small, which has a similar capability score. Early community comparisons put Super VL ahead of Qwen3.8-Flash-Next on throughput and token efficiency, though behind it on accuracy.

Source: @ItsmeAjayKV

Anchor comparison: Nemotron 3.5 Super VL trades peak accuracy for speed. It scores 51.78% on SWE-bench Pro versus 60.37% for Qwen3.8 Flash-Next, but NVIDIA reports it finishes 10,000-task batches up to 4x faster on 4 B200 GPUs.

What You Can Do With Nemotron 3.5 Super VL

NVIDIA lists five target workloads:

1. Coding. Test and validate code, analyze errors and recommend fixes.
2. Web research. Search across sites, connect evidence and verify answers.
3. Document intelligence. Extract data from tables, charts and page layouts.
4. Cybersecurity. Find vulnerabilities and triage threats.
5. Video understanding. Search, content review and incident analysis.

NVIDIA's launch walkthrough video shows the model working out why an AI bill rose 30%, using a dashboard screenshot and a scanned pricing sheet.

Post-training is the main enterprise use. NVIDIA reports these partner results:

• Scale AI's fine-tune for Skydance's vendor-credit agent is expected to cut inference cost by 40x and latency by 5x.
• Ramp expects up to 50% higher throughput and 10x lower latency when converting receipt images to JSON.
• AI21 and Alice improved threat-intelligence accuracy by 41%.

VentureBeat reports that post-trained variants beat Claude Opus 5 and GPT-6.1 Sol on specialized task benchmarks. These partner numbers are vendor-reported and have not been independently reproduced.

Independent tests show limits as well. In a Plotly data-analysis test summarized here, the model answered 30 of 43 questions correctly with a 12-second median time per question. It missed all 3 data-quality questions and all 5 hidden-pattern questions. On KiloBench, an early build scored 53%, the same as GLM 5.2.

Source: @kilocode

How Nemotron 3.5 Super VL Compares

Model	Total / active params	AA Intelligence Index	Notes
Nemotron 3.5 Super VL	120B / 12B	26	196 tok/s, $0.25 per AA task
Nemotron 3 Ultra	550B / 55B	23	NVIDIA's text-only reasoning flagship
Nemotron 3 Super	120B (text-only)	13	Predecessor, March 2026
Qwen3.8 Flash-Next	125B + 51B n-gram	Not yet confirmed	60.37% SWE-bench Pro vs 51.78%
Muse Spark Open (Meta)	816B / 42B	39	Took the top US open-weight spot hours later
MiMo-V2.6-Pro	Not yet confirmed	46	Top open-weight model on the index

Super VL's #1 ranking among US open-weight models lasted only a few hours, until Meta released Muse Spark Open. For context on how the Artificial Analysis index separates frontier models, see the analysis of Muse Wonder's climb to #2. Super VL's advantage is cost and speed. At $0.25 per Intelligence Index task, only MiMo-V2.6-Pro ($0.13) is cheaper among the models Artificial Analysis compares against it. The model is also verbose: it used 170M output tokens on the index, compared with a median of 100M.

Anchor verdict: Nemotron 3.5 Super VL is not the smartest open model, but it is probably the most throughput-efficient open vision agent at 120B scale. That matters most for teams running high-volume, always-on agent loops on their own hardware.

Availability: How to Access Nemotron 3.5 Super VL

The official route is NVIDIA's own stack:

• Weights. BF16 and NVFP4 checkpoints are published under the NVIDIA organization on Hugging Face (NVIDIA-Nemotron-3.5-Super-VL-120B-A12B).
• Serving. NVIDIA NIM, plus day-0 support in vLLM and upstream llama.cpp.
• Post-training. NVIDIA NeMo is the official path. ModelScope's ms-swift also supports inference and fine-tuning.
• Local. NVIDIA's recipe for one DGX Spark uses vLLM with MTP at 1M context. Unsloth's GGUFs range from 46.1GB (1-bit) to 132GB (8-bit). The 4-bit build is 86.9GB, with about 128GB of memory recommended.

Many hosted inference providers added the model on launch day. Several coding agents offered it free for a limited time, and Perplexity added it for Pro and Max subscribers and in its Agent API. If you only need a hosted open-weight agent model with a similar cost profile, DeepSeek-V4.1-Flash is a comparable option.

What We Don't Know Yet

• Memory footprint. Sources disagree on how much RAM local use needs: about 70GB (LM Studio), 82GB (Unsloth's announcement) and 75.1GB for the NVFP4 checkpoint.
• Video in third-party listings. NVIDIA documents video input, but Artificial Analysis lists only text and image input.
• Audio. A pre-release leak says there is no audio encoder. This is unconfirmed by NVIDIA, though no source contradicts it.
• Independent verification. The "up to 5x faster" and post-training claims come from NVIDIA and its launch partners.
• Free-tier duration. Platforms offering free access have not said how long it lasts.
• NVIDIA API list price. Not yet published.

Frequently Asked Questions

Is Nemotron 3.5 Super VL open source?

It is open-weight. NVIDIA publishes BF16 and NVFP4 checkpoints on Hugging Face, and Artificial Analysis lists the license as OpenMDW-1.1. You can download, post-train and self-host it. Whether NVIDIA will release full training data and recipes has not been confirmed.

How much does Nemotron 3.5 Super VL cost?

The weights are free. Hosted pricing tracked by Artificial Analysis is about $0.20 per 1M input tokens and $0.575 per 1M output tokens, which averages $0.25 per Intelligence Index task. Several coding-agent tools offered it free for an unspecified period after launch.

Can I run Nemotron 3.5 Super VL locally?

Yes, on high-memory hardware. NVIDIA publishes an official recipe for one 128GB DGX Spark using vLLM and MTP. Community GGUFs run from 46.1GB (1-bit) to 132GB (8-bit), and upstream llama.cpp supports the architecture.

Nemotron 3.5 Super VL vs Nemotron 3 Ultra: which is better?

Super VL scores higher on the Artificial Analysis Intelligence Index (26 vs 23) with about one-fifth of the parameters. NVIDIA still positions Ultra for the hardest text-only planning work. Super VL is the better fit for visual, document and tool-heavy agent steps.

Does Nemotron 3.5 Super VL support video and audio?

It supports video but not audio. NVIDIA lists text, images, documents, GUIs and video as inputs and text as the only output. A pre-release leak said the model has no audio encoder.

What is the context window of Nemotron 3.5 Super VL?

The context window is up to 1M tokens. Local guides suggest starting at 16K and raising it as memory allows, because long contexts and video need extra memory. For another 1M-token agent model, see this breakdown of GPT-6.1 Sol.

What to Watch Next

Three signals will shape this page's next update:

1. Independent DGX Spark and multi-GPU throughput results from community benchmarkers. The first runs were still in progress at launch.
2. Whether Artificial Analysis adds video input to its listing and confirms official pricing.
3. The October 23 release of Bolt E4B 4GB, which will complete NVIDIA's Super VL, Bolt and Lightning lineup for routing agent steps by tier.

Try related models on Kie.ai: Gemini 3.8 Flash, Claude Sonnet 5.5 and Gemini 4 Argon.