Nvidia Launches Nemotron 3.5 Super VL and Bolt Edge Models for Agents
Nvidia releases three open-weight AI models designed for speed and efficiency, challenging competitors in the agentic AI market.
Key points
- Nemotron 3.5 Super VL is a 120-billion-parameter model optimized for agentic work.
- Two new Bolt models target edge devices with minimal memory requirements.
- Nvidia claims the Super VL is five times faster than nearest competitors.
Nvidia released three new open-weight artificial intelligence models on October 15, including the Nemotron 3.5 Super VL and two Bolt variants. The launch emphasizes speed and efficiency for automated agents, ranging from data center servers to local desktop devices, according to the company’s announcement.
The flagship Nemotron 3.5 Super VL contains 120 billion parameters and targets specialized, always-on agents. Nvidia stated the model delivers the fastest task completion and highest accuracy in its class. This release follows earlier reporting by Threat Times on October 15 regarding the initial announcement of these models.
Performance and Benchmarking
Artificial Analysis benchmarked the Super VL against similarly sized competitors, finding it five times faster than the next best option. InclusionAI’s Ling 3.0 Flash and DeepSeek’s V4.1 Flash placed second and third, respectively, with significant performance gaps. Nvidia said this speed helps enterprise agents complete high-volume tasks across multi-day workflows.
The base model does not top proprietary frontier models without additional training. However, Nvidia demonstrated that post-training can yield substantial gains, citing a 44 percent improvement in medical insurance claims accuracy. This post-trained performance placed the model far above Claude Opus 5 in that specific test case.
Edge and Infrastructure Benefits
Nvidia also introduced Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB for edge computing. The larger Bolt model supports local agentic use, while the smaller 4GB variant handles fast, conversational agents. Both operate locally and off-grid, delivering the highest intelligence per gigabyte of memory relative to other Nvidia models.
Open-weight models allow businesses to post-train on their own data and run infrastructure internally. Nvidia highlighted this as a key advantage over proprietary models, offering greater personalization and security. This release joins a recent surge of Western-made open models challenging China-based labs that previously led the competitive open-weight AI space.
What should you do?
Security and technology professionals should evaluate these models for internal agentic workflows if local processing is a priority. Organizations should benchmark the Super VL against current proprietary solutions for specific high-volume tasks. Teams should also assess the security implications of running open-weight models on local infrastructure versus cloud-based proprietary services.
Background: Edge computing
Edge computing processes data near its source instead of sending it to a distant data center. This approach lowers delay, saves bandwidth, and keeps sensitive information local. It shifts security responsibilities from the core network to individual devices.
Read the full guide: Edge Computing Explained: Why Processing Data Locally Changes Everything
What to do and how to stay safe: Nvidia
- Review internal policies regarding the deployment of open-weight AI models on local infrastructure.
- Benchmark Nemotron 3.5 Super VL against existing solutions for specific agentic tasks to verify speed claims.
- Assess data privacy requirements to determine if local post-training offers sufficient security benefits over cloud APIs.
- Monitor for updates on post-training techniques that may improve base model accuracy for specialized verticals.
Step-by-step guide: Social Engineering Warning Signs and How to Respond
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does Nemotron 3.5 Super VL compare to competitors in speed?
Artificial Analysis benchmarks show it is five times faster than the next best open agentic model in its class.
What are the new Bolt models designed for?
The Bolt models are designed for edge AI, operating locally and off-grid with high intelligence per gigabyte of memory.
Does the base model outperform proprietary models?
No, the base model does not top proprietary frontier models, but post-training can yield significant accuracy gains.
