Nvidia announced three open Nemotron 3.5 models on 15 October 2026, releasing two of them that day, according to its blog post. One of two new edge models, Bolt 74B-A9B 24GB, uses a native 2-bit quantization to fit in 24GB of memory and run agentic tasks locally on NVIDIA RTX and DGX Spark machines.
The headline model, Nemotron 3.5 Super VL, is an open 120-billion-parameter model for always-on agents, Nvidia says. Super VL and Bolt 74B-A9B are available now, and a third model, Bolt E4B 4GB, is due on 23 October.
Business Pill · QUANTIZATION
A one-minute explainer of quantization: how a large AI model is shrunk to fit in less memory. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, the release puts Nvidia’s open models on every tier of Nvidia’s own hardware: the post says the three models cover agent deployments from large enterprise servers down to edge devices and local hardware.
What happened next (30 Oct): the technical report. On 28 October NVIDIA published its technical report on Nemotron 3.5 Super and Nemotron 3.5 Lightning. It describes Super as a 120B-parameter, 12B-active multimodal model and Lightning as a 30B-parameter, 3B-active text model.
NVIDIA reports that the post-trained Super model “achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters.” It also reports that NVFP4 quantization “achieves 99.62% median accuracy recovery relative to BF16”, and that inference optimizations enable “up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.”
Nvidia describes Super VL as delivering “the fastest task completion and highest accuracy for specialized always-on agents.” It says the model handles instruction following, tool use, multistep business workflows and visual understanding.
According to the post, Super VL is comparable with Nvidia’s largest reasoning model at roughly one-fifth the parameter count. It was built in collaboration with the Nemotron Coalition.
Bolt 74B-A9B 24GB is optimized to balance intelligence and speed per GB for local agentic workloads, the post says. Bolt E4B 4GB is optimized for intelligence per GB for fast conversational agents, on GeForce RTX, Jetson and RTX/DGX Spark.
Nvidia says the three join Lightning, released in August for high-throughput workloads, and Ultra, for the hardest reasoning tasks. In its words, the three new models “together cover the full range of agent deployments, from large enterprise servers down to edge devices and local hardware.”
Nvidia says Nemotron 3.5 Bolt 74B-A9B uses a native 2-bit quantization to fit in 24GB of memory; Super VL is an open 120-billion-parameter model and Bolt E4B 4GB is due on 23 October (Nvidia blog, 15 October 2026).
What Was Posted on X
Today, weโre expanding our Nemotron family of open models with Nemotron 3.5 Super VL.ย
Bringing visual understanding, reasoning, and tool use together to help AI agents complete more work in less time.ย
Open models โก๏ธ more control to customize, deploy, and build on your own terms.
Nvidia’s main account announced the Nemotron 3.5 Super VL expansion on X on 15 October 2026, quoting the NVIDIA AI account’s launch post, which gives the model’s size as “120B total parameters with 12B active” and describes it as combining reasoning and tool use with text, image, document and video understanding for long-running AI agents.
Great report from NVIDIA.
They report that Nemotron 3.5 Super reaches 51.78% on SWE-bench Pro. That beats Nemotron 3 Ultra with about 4.6x fewer active parameters.
Super has 120B parameters with 12B active, and it handles text, images, and video. Lightning is a 30B text model with 3B active.
The technical report covers every stage of training. Super gets mid-training on agentic software engineering, then supervised fine-tuning, RL, and multi-teacher on-policy distillation.
Serving numbers are also included. The stack reaches up to 4,856 output tokens/s/GPU on GB200 while holding at least 50 tokens/s/user. NVFP4 quantization recovers 99.62% median accuracy relative to BF16.
DAIR.AI highlighted NVIDIA’s technical report on Nemotron 3.5 Super and Lightning, citing 51.78% on SWE-bench Pro for Super and about 4.6x fewer active parameters than Nemotron 3 Ultra.
What Nvidia Claims on Speed
Nvidia calls Super VL “the fastest open agentic model in its class” and says it is “5x faster than the next best.” The post says that on the Artificial Analysis Intelligence Index leaderboard it achieves up to 5x faster throughput compared to other models in its class.
Running the model on NVIDIA Dynamo, the post says, organizations can serve up to 8x more users with 85% lower latency. These are Nvidia’s own figures; the post does not publish the test setup in the text.
For Bolt 74B-A9B, Nvidia says it is the #1 open model in its class on PinchBench and the #2 open model for BrowseComp.
Who Is Customizing Super VL
Nvidia says dozens of organizations are post-training Super VL on their own data, tools, policies and workflows. The post says Scale AI post-trained the model for Skydance’s vendor credit finance agent and for EY’s M&A due diligence agent, and AfterQuery customized it for Hebbia for credit agreement analysis.
Software companies integrating the model, per Nvidia, include Atlassian, CrowdStrike, Dassault Systèmes, Perplexity and Salesforce, and Synopsys software for semiconductor design verification interoperates with it.
Open Weights on Nvidia Hardware
The model ships with open weights, open datasets and training recipes, the post says. Nvidia says Super VL was trained on about 26 trillion tokens and post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments.
Nvidia says Super VL runs across its full hardware stack: DGX Spark, DGX Station, RTX PRO and RTX systems at the desktop, Jetson at the edge, and data centre and cloud environments. It is available on Hugging Face, Open Router, ModelScope and as NVIDIA NIM.
Bolt 74B-A9B is on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM, the post says, and is built for NVIDIA RTX Spark laptops arriving on 16 October.
As we read it, the 24GB figure is the point of Bolt: the post says native 2-bit quantization lets a 74B-A9B model fit in 24GB of memory and run locally on RTX and DGX Spark.
As we read it, the customization list is the demand signal Nvidia chose to show: the post names Scale AI’s agents for Skydance and EY, Hebbia’s credit agreement analysis and Ramp’s financial document intelligence.
As we read it, the openness runs on Nvidia’s terms: the post says the model ships with open weights, datasets and training recipes, and that it runs across Nvidia’s hardware stack and as NVIDIA NIM.
Nvidia, in its 15 October 2026 blog post
“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.”
Three Implications
ENTERPRISES Nvidia says organizations can post-train Super VL on their own data and run it wherever their security and data-residency requirements dictate.
DEVELOPERS The post says Bolt 74B-A9B is on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
PC MAKERS Nvidia says Bolt 74B-A9B is built for NVIDIA RTX Spark laptops arriving on 16 October and also runs on GeForce RTX PCs.
The framework sets out the pattern as: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize).”
As we read it, Nvidia’s release follows that shape: the models, weights and datasets are open, and the post says they run across Nvidia’s own hardware; the framework does not discuss this release.
Lineup from Nvidia’s 15 October 2026 blog post: Super VL and Bolt 74B-A9B are available now and Bolt E4B is due on 23 October; Lightning was released in August.
What Is Not Established
Nvidia’s post gives no price for NIM or for enterprise support, and no usage or download figures. The speed and latency figures are Nvidia’s, and the post does not name the models it compared against.
Bolt E4B 4GB is announced for 23 October and is not yet available, according to the post.
Nvidia announced open Nemotron 3.5 models on 15 October 2026, from a 120-billion-parameter Super VL for enterprise agents to a 74B-A9B Bolt model that fits in 24GB on its RTX and DGX Spark machines.
Nvidia says Super VL runs across its hardware stack, from data centre to desktop to Jetson, and that Bolt 74B-A9B runs on RTX PCs, DGX Spark, DGX Station and RTX PRO workstations.
A note on sourcing. We read Nvidia’s 15 October 2026 blog post on Nemotron 3.5 Super VL and Bolt in full; every model specification, performance claim and customer name comes from it, and the speed and latency figures are Nvidia’s own. The Jensen Huang and Ramp context comes from our own earlier reporting. We did not use other outlets’ reports.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
