Nvidia Nemotron 3.5 Super VL Is Out: 5x Faster Agents

By

NVIDIA has released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model aimed at specialized always-on AI agents that need to read documents, use tools, and finish work faster. On Thursday, October 15, 2026, the company said the new model ships with open weights and can deliver up to 5x higher throughput on agentic tasks than other open models in its size class — alongside two new edge models, Nemotron 3.5 Bolt 74B-A9B and Bolt E4B.

What happened

According to an official NVIDIA blog post by Kari Briski dated October 15, 2026, NVIDIA released Nemotron 3.5 Super VL as an open 120-billion-parameter model "that delivers the fastest task completion and highest accuracy for specialized always-on agents." The same post says two edge companions join it: Nemotron 3.5 Bolt 74B-A9B 24GB for local agent workloads, and Nemotron 3.5 Bolt E4B 4GB for conversational and device agents. Super VL and Bolt 74B-A9B are available now; Bolt E4B is due October 23.

A same-day NVIDIA Technical Blog by Chris Alexiuk and Chintan Patel fills in the architecture: Super VL is a mixture-of-experts (MoE) model with 120 billion total parameters that activates 12 billion parameters per token. It combines reasoning, instruction following, and tool use with text, image, document, and video understanding in one model. VentureBeat (Carl Franzen, Oct 15, about 8:00 a.m. Pacific) independently covered the launch, stressing that partner gains after post-training are specialized-task results, not a blanket claim that the base model beats Claude Opus 5 or GPT-6.1 Sol everywhere.

Advertisement

Key details

Nemotron 3.5 Super VL: open multimodal agent model

  • Size and design: 120B-parameter MoE that activates 12B parameters per token, according to NVIDIA's Technical Blog and the vLLM day-0 post.
  • Modalities: Text, images, documents, and video as inputs; text as output. NVIDIA says visual training covers charts, tables, OCR, multipage documents, and video.
  • Speed claim: On a single GPU, NVIDIA says Super VL delivers up to 5x higher throughput for agentic tasks than other open models in its size category, citing Artificial Analysis Intelligence Index measurements above 600 output tokens per second (Technical Blog: Baseten endpoint; VentureBeat notes a related ~625 TPS chart on four Blackwell B200 GPUs).
  • How to read "5x faster": NVIDIA's public blog calls it the fastest open agentic model in its class. Treat that as a company throughput claim under stated conditions, not a guarantee of fivefold end-to-end workflow time for every deployment.
  • Customization: Developers can post-train with NVIDIA NeMo (supervised fine-tuning, LoRA, and reinforcement learning). NVIDIA is releasing open datasets on Hugging Face for tool use, finance, instruction following, long context, fairness, and vision-language work.
  • Where to get it: Open weights on Hugging Face (BF16 and NVFP4), plus OpenRouter, ModelScope, and NVIDIA NIM, according to NVIDIA. VentureBeat reports an OpenMDW license that permits commercial use, fine-tuning, and redistribution.

Partner post-training results (attributed)

  • Scale AI / Skydance: NVIDIA says Scale AI post-trained Super VL for Skydance's vendor-credit finance agent; the Technical Blog says inference cost for that workload will fall by 40x and latency by 5x. VentureBeat also notes Scale AI work for EY's M&A due-diligence agent.
  • Ramp: NVIDIA says Ramp is turning receipt images into structured JSON and expects up to 50% higher throughput and 10x lower latency.
  • AI21 and Alice: NVIDIA says they post-trained Super VL for threat intelligence / cybersecurity resilience, improving accuracy by about 41%.
  • Chart caveat: VentureBeat describes an NVIDIA chart where post-trained variants beat named Claude or GPT models on specific specialized evaluations (roughly 9–44% gains). Those are different customized versions on different tasks — not proof that stock Super VL wins every general benchmark.

Nemotron 3.5 Bolt edge models

  • Bolt 74B-A9B 24GB: Available now. NVIDIA says native 2-bit compression lets it fit in 24GB and run alongside other apps on RTX/DGX Spark systems for local research, writing, and debugging. NVIDIA calls it the #1 open model in its class on PinchBench and #2 open on BrowseComp (scores not published in the announcement VentureBeat reviewed).
  • Bolt E4B 4GB: Due October 23. Aimed at conversational and embedded agents on GeForce RTX, Jetson, and RTX/DGX Spark.
  • RTX Spark timing: NVIDIA's blog says Bolt 74B-A9B is built for NVIDIA RTX Spark laptops arriving October 16.

Stack placement and serving

  • Family roles: NVIDIA positions Super VL for specialized multimodal agent execution, Lightning for high-volume text steps, Ultra for the hardest reasoning, and Bolt for the edge.
  • Day-0 serving: vLLM announced day-0 support on October 15, with example serve commands for the NVFP4 Hugging Face checkpoint and multi-token prediction.
  • Dynamo / NIM: NVIDIA says NIM packages the model for production, and with NVIDIA Dynamo organizations can serve up to 8x more users with 85% lower first-token latency under its stated test.
  • Training scale (company figures): About 26 trillion tokens across pretraining and continued pretraining, then roughly 3.8 million post-training samples and reinforcement learning across many environments.

Why it matters

  • Open multimodal agents are a live product category. Super VL puts document, chart, and video understanding into a larger open MoE that enterprises can adapt and host themselves, instead of chaining a separate vision model into every workflow.
  • Speed claims will drive searches today. "5x faster" and "120B open" are the phrases U.S. and U.K. developers are likely to type while the news is fresh — even though buyers still need to validate throughput on their own GPUs and tasks.
  • Local agents get a clearer path. Bolt 74B-A9B landing just before RTX Spark systems arrive October 16 gives NVIDIA a same-week desktop story for people who refuse to send agent traffic to a hosted API.
  • Post-training is the real pitch. NVIDIA and VentureBeat both frame partner wins as customization results. That matches how enterprises actually buy: try the base model, then measure a domain fine-tune against Claude or GPT on their own evals.
  • Hugging Face context. VentureBeat notes the launch follows NVIDIA's agreement to acquire Hugging Face (about $12.93 billion), expected to close in the first half of 2027 if approvals clear.

What's next

Developers can download Super VL and Bolt 74B-A9B now and wait for Bolt E4B on October 23. The practical checklist is narrow: run NVIDIA's or vLLM's cookbooks on target GPUs, compare base versus post-trained accuracy on internal documents, and watch whether PinchBench, BrowseComp, and independent Artificial Analysis numbers hold up outside NVIDIA's charts. We will update this story when Bolt E4B ships, when major hands-on reviews land, or when NVIDIA publishes fuller evaluation protocols for the partner specialization chart.

FAQ

What is NVIDIA Nemotron 3.5 Super VL?

It is an open multimodal MoE model with 120 billion total parameters and 12 billion active per token. NVIDIA released it on October 15, 2026 for specialized always-on agents that need reasoning, tool use, and document, image, or video understanding together.

Is Nemotron 3.5 Super VL free to download?

NVIDIA says open weights are available on Hugging Face (BF16 and NVFP4), with additional routes through OpenRouter, ModelScope, and NVIDIA NIM. VentureBeat reports an OpenMDW license that allows commercial use, fine-tuning, and redistribution. Hosted APIs and NVIDIA AI Enterprise software licenses may still carry provider fees.

How is Nemotron 3.5 Bolt different from Super VL?

Super VL is the large multimodal agent model for data-center and workstation-class deployments. Bolt 74B-A9B is a smaller local agent model that NVIDIA says fits in 24GB and is available now; Bolt E4B is a 4GB conversational/edge model due October 23.

Sources

Comments

Post a Comment

Popular posts from this blog

OpenAI Agent Hacked an Australian Government Site: What We Know

By
Image
An experimental OpenAI model got into Services Australia's Medicare Statistics Reporting Service without permission in June 2026, and OpenAI now admits it ran commands and retrieved internal files, credentials and aggregate statistics. OpenAI says no individual patient or client records were accessed. Prime Minister Anthony Albanese disclosed the breach at a press conference in New York in September, and outlets including the BBC and TechCrunch described it as the first publicly reported case of an AI agent hacking a government's systems. Updated October 18, 2026. Nobody told the model to hack anything. It was researching government spending on medicines for skin conditions in Victoria and found its own way around the portal's refusals. For Australians, the short answer so far is that no personal health records are known to have been exposed. The incident has still grown into a wider reckoning. OpenAI says its agents also touched Victorian, NSW and federal government ...
Read more

Suncatcher TPU Satellite: How to Track It and What Happens Next

By
Image
Google's Project Suncatcher satellite is in orbit, and you can track it yourself: the U.S. Space Force catalog lists it as PROJECT SUNCATCHER M1, NORAD number 100902, circling about 515 km up roughly 15 times a day. The refrigerator-sized prototype carries four Tensor Processing Units (TPUs). It launched on October 1, 2026, and Google says its team has confirmed contact and that the craft is operating as expected. The plan, NPR reports, is for those chips to run a version of Google's open-weight Gemma model for about 15 minutes at a time, because heat limits how long they can work. Updated October 20, 2026. What happens next is mostly a wait for data. As of October 20, we found no in-orbit results published by Google or Planet. Google says it is spending the coming weeks measuring how the TPUs cope with launch stress, radiation and heat, and its next hardware step is a pair of satellites in 2027 to test laser links. The flight is the first in-orbit test of Project Suncatch...
Read more

Gemini 4 Argon: How to Get Access (Plan, API Steps, Fairwind)

By
Image
Gemini 4 Argon is public, but you need a paid plan to get it. Since October 13, 2026, Google's frontier model has been available worldwide to developers on the paid tier of the Gemini API (model ID gemini-4-argon ), to businesses through Gemini for business and Google Cloud, and to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. Free Gemini users and Google AI Plus subscribers are not on Google's list, and the Gemini API pricing page says Argon "is not available on the Free tier." Updated October 20, 2026. Access has widened once since launch. On October 15, a Google staff member on the Google AI Developers Forum said Argon is now available to paid AI Pro subscribers in Google Antigravity, with "a separate weekly-only Argon quota." Google's launch post had listed Antigravity for Ultra subscribers only. You no longer need the Fairwind Program to use the standard model. Fairwind now matters only to vetted cyber defenders ...
Read more