NVIDIA Launches Nemotron 3.5 Super VL and Bolt Open AI Models
NVIDIA launches open Nemotron 3.5 Super VL and Bolt models, spanning data centers to edge devices with faster agents, local execution and customization.
Summary
On Oct. 15, 2026, NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter agent model developed with the Nemotron Coalition. NVIDIA says it matches its largest reasoning model at about one-fifth the parameters and reaches up to 5x the throughput of class peers on Artificial Analysis Intelligence Index, covering instruction following, tool use, multistep workflows, images, webpages, PDFs, OCR, tables and charts. Training used about 26 trillion tokens, roughly 3.8 million post-training samples and reinforcement learning across dozens of environments. Open weights, datasets and recipes support on-premises customization through NeMo and Python pipelines; NIM receives updates every two weeks, while Dynamo enables up to 8x more users at 85% lower latency.
Bolt 74B-A9B uses native 2-bit quantization to fit 24GB and run local research, writing and debugging on RTX and DGX Spark; it ranks No. 1 on PinchBench and No. 2 among open models on BrowseComp. Bolt E4B fits 4GB for voice assistants, interactive characters and embedded interfaces on GeForce RTX, Jetson and RTX/DGX Spark. Super VL and Bolt 74B-A9B are available now through Hugging Face, Open Router, ModelScope and NIM, while E4B arrives Oct. 23. The 74B model supports llama.cpp, Ollama, LM Studio, Unsloth and vLLM, including RTX Spark laptops arriving Oct. 16. Super VL spans DGX Spark, DGX Station, RTX PRO, RTX, Jetson, cloud and data centers.
Scale AI, Skydance, EY, AfterQuery, Hebbia, AI21, Alice, Mercor, ModelScope, Ramp, Dream, Lila Sciences, R1, TrainLoop and Nolla are customizing Super VL for finance, cybersecurity, coding, science and healthcare. Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys are integrating it for search, security, engineering, government automation and chip verification. The releases extend Nemotron beyond August's high-throughput Lightning and hardest-reasoning Ultra, giving enterprises open models from servers to edge devices without sending local workloads to the cloud.
Positives
- Up to 5x higher throughput than class peers positions Super VL for high-volume, long-running agent workflows.
- Open weights, datasets and training recipes let organizations customize Super VL and retain workloads on their own infrastructure.
- Native 2-bit quantization lets Bolt 74B-A9B fit within 24GB while running alongside other local applications.
- NVIDIA Dynamo can support up to 8x more users with 85% lower latency.
- No. 1 on PinchBench and No. 2 on BrowseComp give Bolt 74B-A9B strong open-model benchmark positions.
- More than 20 named organizations are customizing or integrating Super VL across finance, security, healthcare, science, engineering and government.
Risks & concerns
- Bolt E4B remains unavailable until Oct. 23, eight days after the initial Super VL and Bolt 74B-A9B release.
- Bolt 74B-A9B requires 24GB of memory, while only the smaller E4B targets a 4GB footprint.
- No. 2 on BrowseComp means Bolt 74B-A9B does not lead that open-model benchmark.
- Super VL still contains 120 billion parameters, leaving Bolt as NVIDIA's option for memory-constrained edge deployments.
- Named deployment support centers heavily on NVIDIA DGX, RTX and Jetson hardware.