On October 15, 2026, Nvidia added three new open-weight models to the “Nemotron 3.5” family, led by the 120-billion-parameter multimodal model “SuperVL”. The new collection aims to run artificial intelligence agents with high efficiency across cloud systems, desktops, and laptops.
Three open models led by SuperVL to power specialized agents
Cutting inference costs by 40x and running locally on Spark computers
Nvidia’s open model strategy to compete with Chinese companies
Cloud and edge computing integration in the AI agent ecosystem
Frequently asked questions
Three open models led by SuperVL to power specialized agents
On Thursday, October 15, 2026, Nvidia added three new open-weight models to its “Nemotron” smart model family. Headlining this lineup is the flagship “Nemotron 3.5 SuperVL”, a multimodal model featuring 120 billion software parameters designed specifically to run “specialized agents operating around the clock”.
Accompanying this massive model are two edge computing models: the “Nemotron 3.5 Bolt 74B-A9B” with a 24GB capacity, and the “Nemotron 3.5 Bolt E 4B” with a 4GB capacity. Together, these models aim to power AI agents on everything from massive data center systems to desktop and laptop computers, according to Yahoo Finance. The “SuperVL” and “Bolt 74B” models are available now, while the “Bolt E 4B” model is set for release on October 23.
According to the Nvidia developer blog, the “SuperVL” model relies on a mixture-of-experts design that activates only 12 billion of the 120 billion parameters per token, handling text, images, documents, and video within a single model. The company emphasizes that the model achieves up to 5x higher throughput in agent tasks compared to other open models of similar size when running on a single GPU, and scored 80.8% on the ChartQA visual reasoning benchmark.
Cutting inference costs by 40x and running locally on Spark computers
The “SuperVL” model was distilled from the larger “Nemotron 3 Ultra” model, allowing enterprises to later retrain it on their own private data. Nvidia pointed to early results from several companies: Scale AI expects inference costs for a vendor credit workload to drop by 40x, while Ramp expects to cut financial receipt parsing latency by up to 10x. The model is available via Hugging Face, OpenRouter, and ModelScope, and as a microservice through Nvidia NIM.
On the local hardware front, the “Bolt 74B-A9B” model contains roughly 74 billion parameters and utilizes 9 billion of them per token. Two-bit quantization technology allows it to fit inside just 24GB of memory. Like the rest of the Nemotron models, its weights, training data, and recipes are fully open, running across Llama.cpp, Ollama, LM Studio, Unsloth, and vLLM environments.
Nvidia designed this model specifically for its new “RTX Spark” laptops going on sale Friday, October 16, whereas the “Bolt E 4B” model was built for physical and conversational agents such as voice assistants and interactive gaming characters on GeForce RTX, Jetson, and DGX Spark hardware.
Nvidia’s open model strategy to compete with Chinese companies
Open-weight models have become a central part of Nvidia’s long-term growth strategy. Lower-cost models can attract more enterprises and corporations into the AI world, which in turn will drive future demand for electronic chips, as noted in the Yahoo Finance report.
Chinese development firms, such as Moonshot AI and Alibaba, currently lead the global conversation around open models through the Kimi and Qwen families.
In this context, Nvidia CEO Jensen Huang urged American companies to build and adopt more open models so the sector does not remain trapped within a few closed systems. Meanwhile, ZDNet described the launch of “SuperVL” as “another step forward for American open-weight AI.”
Cloud and edge computing integration in the AI agent ecosystem
The rollout of the three Nemotron 3.5 models demonstrates Nvidia’s comprehensive vision for the future of smart agents, integrating the massive data center capabilities of the SuperVL model with the efficiency of the Bolt 74B and Bolt E 4B models designed to run locally on RTX Spark laptops and Jetson systems.
By making software weights and training data freely available and encouraging American companies to adopt open models against Chinese competition from the Kimi and Qwen families, Nvidia expands the reach of its software while ensuring sustained robust demand for its graphics processors in both the cloud and personal devices.
Frequently asked questions
Question: What are the three models launched by Nvidia within the Nemotron 3.5 family?
Answer: Nvidia launched the 120-billion-parameter SuperVL model for cloud systems, alongside edge models Bolt 74B with 24GB capacity and Bolt E 4B with 4GB capacity for local devices.
Question: What performance does the SuperVL model achieve for companies like Scale AI and Ramp?
Answer: It delivers 5x higher throughput and scored 80.8% on the ChartQA benchmark. Scale AI expects a 40x reduction in inference cost, while Ramp expects a 10x speedup in receipt processing.
Question: How does the Bolt 74B model manage to run on laptops?
Answer: It activates only 9 billion parameters per token and relies on 2-bit quantization to run within 24GB of memory on RTX Spark laptops.
Question: Why does Nvidia and its CEO Jensen Huang support open-weight models?
Answer: To encourage more organizations to adopt AI, boost chip demand, compete with open Chinese models like Kimi and Qwen, and avoid closed monopolies.
