Nvidia released Nemotron 3.5 Super VL on October 15, 2026, a 120-billion-parameter open-weight model built to run AI agents that handle both images and text. The company shipped two smaller Bolt models alongside it, and the Bolt 74B-A9B was available at launch for machines with 24GB of memory. The release landed in the same week AI chip stocks lost roughly $1 trillion in value.
Nvidia laid out the lineup in its launch blog post, describing Super VL as an open model for always-on specialized agents. Startup Fortune reports that the weights appeared on Hugging Face on October 13, two days before Nvidia's announced date. Treat October 15 as the official launch date until the model card settles the question.
What Nvidia released, model by model
Super VL is the headline model. The Decoder reports that it uses a mixture-of-experts design that activates 12 billion of its 120 billion parameters for each token. Nvidia distilled it from Nemotron 3 Ultra, the company's largest reasoning model.
Nvidia says Super VL matches that larger model on multimodal agent tasks at about one-fifth the parameter count, according to Unite.ai and Verdict. The model ships with open weights, open datasets and training recipes. Organizations can post-train it on their own data. Yahoo Finance reports the base model was trained on about 26 trillion tokens, followed by post-training on about 3.8 million samples.
Nvidia built the model with the Nemotron Coalition, which includes Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs. The model is distributed through Hugging Face, ModelScope, OpenRouter and Nvidia's NIM microservice.
Not every outlet agrees on the specs. Startup Fortune reports 121 billion total parameters and 13 billion active per token, which differs from Nvidia's 120 billion and 12 billion. Check the model card before you cite either set of numbers.
The two Bolt models target edge devices and local PCs. The table below summarizes the lineup.
| Model | Design | Runs on | Availability |
|---|---|---|---|
| Nemotron 3.5 Super VL | 120B total, 12B active per token | DGX Spark, DGX Station, RTX, RTX PRO and Jetson systems | Released October 15 |
| Bolt 74B-A9B | Native 2-bit weights | 24GB of memory, built for RTX Spark laptops | Available at launch |
| Bolt E4B | Smaller edge model | 4GB of memory | Scheduled for October 23 |
Which machines can run the new models
The Bolt 74B-A9B is the most practical release for individual buyers. According to CellCog, it uses native 2-bit weights to fit in 24GB of memory. Nvidia said the RTX Spark laptops it is built for would arrive on October 16.
Super VL is a different proposition. A 120-billion-parameter model still has to fit in memory, even though only 12 billion parameters activate per token. The active-parameter count should speed up each response. It will not shrink the memory requirement.
Nvidia claims up to 5x higher throughput than comparable open models on a single GPU, as reported by SQ Magazine. Those performance figures come from Nvidia itself and have not been independently verified. Treat them as a starting point for your own testing.
The Decoder also notes that getting the most out of Super VL means buying Nvidia hardware. The release is framed around DGX, RTX and Jetson systems, so the strongest benefits go to people already invested in Nvidia's stack.
How the release fits the open-model race
Meta released its own large open-weight model the same week. According to TechCrunch, Muse Spark Open has 816 billion parameters, a 1 million token context window and can run on Macs.
Mistral's October 6 flagship set the tone earlier in the month. CNBC reported that Mistral said the model rivals the best open systems from China. Open weights have become a central front in the contest between US and Chinese AI labs.
Nvidia's incentive is more direct than Meta's or Mistral's. A model tuned for its own chips makes its GPUs more useful, so every open release also supports hardware sales. Startup Fortune describes Nvidia's open-model effort as a $26 billion, five-year bet against China's AI lead.
For a sense of how quickly the field moves, see our coverage of DeepSeek's V4.1-Flash-1017 release, which also shipped with open weights and unchanged prices. Whether Nvidia's models will pull developers away from Chinese open models is still unclear.
Why the chip selloff makes this release harder to read
AI Weekly reports that about $1 trillion was lost across 20 major AI and semiconductor stocks in the week of October 15. SK Hynix fell more than 10% in Seoul on Thursday, October 15. Nvidia sits at the center of that trade, so investors are weighing its open-model strategy against a sharp repricing of AI hardware.
The timing does not prove the model launch caused the drop. Markets were moving on many factors that week. The more useful question is whether open models shift spending between cloud APIs and local hardware. That shift would favor chipmakers only if buyers keep purchasing GPUs.
The financing side of AI infrastructure is also under strain, as our report on the CoreWeave bond selloff shows.
Who gains and who should be cautious
Developers and small companies stand to gain the most from downloadable weights. Nvidia says dozens of organizations are already customizing Super VL for financial document analysis, medical claims and scientific discovery. Running models locally may lower API bills, but it replaces those fees with hardware costs and maintenance.
PC buyers face a different calculation. A 24GB machine is a meaningful purchase, and the Bolt 74B-A9B only makes sense if you plan to run local agents regularly. Casual users will get more value from cloud tools unless privacy or offline use matters to them.
Reflection AI's Beam model also ships under a permissive license, as our Beam coverage explains. Compare license terms across open releases before you build a product on any of them.
What to do before you download
- Read the license. One secondary source reports the OpenMDW-1.1 license, which requires keeping license notices. Confirm the terms on the model card before any commercial use.
- Check your memory. The Bolt 74B-A9B needs 24GB of memory, so confirm your machine meets that before starting a large download.
- Verify the specs. Sources disagree on the parameter counts and the release date, so use the numbers in the official model card.
- Wait for Bolt E4B if you need an edge model. Nvidia scheduled the 4GB model for October 23, and it is not yet available.
- Test on your own workload. Nvidia's throughput claims are its own, so measure speed and accuracy on the tasks you care about.
What to watch next
The October 23 release of Bolt E4B is the next concrete milestone. Independent benchmarks of Super VL against DeepSeek and other open models will show whether Nvidia's one-fifth parameter claim holds up outside its own tests. The license terms will also need confirmation from primary documentation.
Over the next few weeks, watch whether developers adopt the Nvidia models in production or only test them. Watch the chip sector too, because the same investors who sold AI hardware last week will judge whether open models strengthen or erode demand for GPUs.
