Open From Data Center to Desktop: NVIDIA Expands Nemotron Family

The section lengths are in range. I'll add the two missing words and emit the article.NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, an open 120-billion-parameter model for the fastest task completion and the highest accuracy on specialized always-on agents, according to NVIDIA Blog's report. Nemotron 3.5 Bolt 74B-A9B 24GB joins it for local agentic workloads, balancing intelligence and speed per GB. Nemotron 3.5 Bolt E4B 4GB is tuned for intelligence per GB on fast conversational agents. Super VL and Bolt 74B-A9B are available now. Bolt E4B comes Oct. 23.

Teams running agents on NVIDIA machines get an open model per tier, from enterprise servers to edge devices and local hardware. Super VL and Bolt join Lightning, released in August for high-throughput work, and Ultra, for the hardest reasoning tasks. This piece covers sizes, speed, training, post-training work, install paths, and later dates.

What Open From Data Center to Desktop shipped

Nemotron 3.5 Super VL, built with the Nemotron Coalition, targets leading accuracy on instruction following, tool use, multistep business workflows, and visual understanding. NVIDIA says that matches its largest reasoning model at roughly one-fifth the parameter count: frontier-level intelligence for high-volume agentic workflows. It is the fastest open agentic model in its class, 5x faster than the next best, including workflows that span days or weeks. On the Artificial Analysis Intelligence Index leaderboard, throughput is up to 5x other models in its class.

Bolt 74B-A9B 24GB runs on NVIDIA RTX/DGX Spark, fitting in 24GB beside other applications via native 2-bit quantization. NVIDIA says it is "smart enough to get most jobs done, yet much faster than the smartest models." Local harnesses research the web, write, and debug code without the cloud. It is #1 on PinchBench and #2 on BrowseComp among open models in its class. Bolt E4B 4GB is built for physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark, with natural multi-turn conversation for voice assistants, interactive characters, and embedded interfaces.

What changed for builders in Open From Data Center to Desktop

Organizations can post-train, deploy, and oversee Super VL on their own data and infrastructure, unlike closed frontier models, and it ships open weights, open datasets, and training recipes. About 26 trillion tokens covered pretraining and continued pretraining, then about 3.8 million samples and reinforcement learning across dozens of environments. NVIDIA NeMo, Python pipelines, and NeMo Gym customize it. Hugging Face datasets cover agentic tool use, finance, instruction following, long context, fairness, and vision-language understanding. On NVIDIA Dynamo, NVIDIA states up to 8x more users and 85% lower latency.

Dozens of organizations post-train Super VL on their data, tools, policies, and workflows: Scale AI for Skydance's vendor credit finance agent and EY's M&A due diligence agent; AfterQuery for Hebbia credit agreement analysis; AI21 and Alice for cybersecurity attack resilience; Mercor on APEX-SWE for agents that plan, write, and debug code; ModelScope ms-swift for specialized multimodal applications; Ramp for financial document intelligence; Dream for sovereign AI-powered cybersecurity research and vulnerability analysis; Lila Sciences for scientific discovery and search; R1 for medical claims and revenue cycle management. TrainLoop used Nolla dermatology data. NVIDIA says Nolla is the first company approved for AI to write initial patient prescriptions.

How to install or upgrade Open From Data Center to Desktop

Super VL is on Hugging Face, Open Router, the ModelScope community, and NVIDIA NIM now. It runs on DGX Spark, DGX Station, RTX PRO, and RTX desktops, on Jetson at the edge, and in data center and cloud. Bolt 74B-A9B 24GB is on Hugging Face, with llama.cpp, Ollama, LM Studio, Unsloth, and vLLM. It is built for RTX Spark laptops arriving Oct. 16 and also runs on GeForce RTX PCs, DGX Spark, DGX Station, and RTX PRO workstations. Bolt E4B 4GB is dated Oct. 23.

The Super VL open-weights id is nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4. Install or update vLLM, then serve that id to pin Super VL as the server default. Bolt's Ollama and llama.cpp id is nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF. The vLLM id is nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB. LM Studio lists nvidia/nemotron-3.5-bolt-74b-a9b. The id you launch is the session model. This release publishes no slash command and no extra default flag. Post-training uses NVIDIA NeMo, Python pipelines, and ms-swift. The commands below now pin Super VL, the GGUF Bolt, and the vLLM Bolt.

pip install vllm

vllm serve "nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4"

ollama run hf.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF

vllm serve "nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB"

llama cli -hf nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF

Gotchas and compatibility in Open From Data Center to Desktop

The three sizes are not substitutes. Super VL is 120 billion parameters for high-volume multimodal agents. Bolt 74B-A9B is the 24GB 2-bit model that runs beside other apps. Bolt E4B is the 4GB conversational and physical model, and it is not out yet. Ultra still covers the hardest reasoning. Lightning, from August, still covers high-throughput work. One-fifth the size compares Super VL with NVIDIA's largest reasoning model on those agentic tasks, not with every Ultra score. The 5x figures are in-class speed, including Artificial Analysis Intelligence Index throughput.

Super VL vision covers optical character recognition, tables, charts, and structured information in visual documents, web pages, PDFs, and images. It extracts text, reads graphs and charts, and returns insights. One model carries context from document to data to decision. Teams run the weights where security and data-residency rules require and can adapt them. Insurance claims can be scanned for fraud markers on repeat. Credit scores can be checked against a lender's own criteria. NVIDIA says, "The model stays on premises. The data stays private."

What to watch after Open From Data Center to Desktop

Atlassian is integrating Super VL for agentic search across Jira and Confluence. CrowdStrike is customizing it for agentic blue teaming in enterprise security. Dassault Systèmes is using it for engineering design and simulation. Perplexity is using it for search and agentic tasks in Perplexity Search and Computer. Salesforce is customizing Nemotron so agents digitize and automate complex government workflows in secure customer environments through Missionforce Operations. Synopsys software for semiconductor design verification also interoperates with that same open model today.

RTX Spark laptops built for Bolt 74B-A9B arrive Oct. 16. Those 24GB weights already run on GeForce RTX PCs, DGX Spark, DGX Station, and RTX PRO workstations. Bolt E4B 4GB comes Oct. 23 for GeForce RTX, Jetson, and RTX/DGX Spark and is not yet in today's downloads. NVIDIA NIM, a production-ready microservice with performance-optimized, validated configurations and enterprise-grade support, refreshes every two weeks with security patches and performance improvements. Hugging Face datasets and NeMo Gym support further Super VL post-training.