Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models

NVIDIA has released Nemotron 3.5 Super VL, an open 120-billion-parameter vision-language model aimed at specialized always-on agents, together with two edge models in a new Bolt line: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. Super VL and Bolt 74B-A9B are available now; Bolt E4B follows on October 23.
The three models are positioned to span agent deployments from enterprise servers down to local hardware and edge devices. They slot into an existing Nemotron family alongside Lightning, released in August for high-throughput workloads, and Ultra, which handles the hardest reasoning tasks. Bolt is the line's edge entry. NVIDIA frames the pitch as removing a trade-off between capability and deployability for organizations running agents on its infrastructure.
Super VL was built with the Nemotron Coalition and is claimed to reach leading accuracy on multimodal agentic tasks — instruction following, tool use, multistep business workflows and visual understanding — comparable to NVIDIA's largest reasoning model at about one-fifth the parameter count. NVIDIA describes it as the fastest open agentic model in its class, at 5x faster than the next best, and cites the Artificial Analysis Intelligence Index leaderboard for throughput figures of up to 5x versus other models in the class. Those are vendor-stated comparisons, and the relevant point for buyers is that the model is meant for sustained, high-volume agent work rather than one-off hard reasoning.
On the vision side, the model targets what agents actually meet: optical character recognition, tables, charts and structured information in visual documents. It can take a webpage, pull out text, read graphs and charts and return findings, with the same applying to PDFs and other documents. Keeping text and visual understanding in one model means an agent does not have to hand tasks between separate systems, so context carries from document to data to decision. In practice that suits document-heavy pipelines — extracting fields from contracts, reading financial statements, pulling figures out of reports — where routing between an OCR tool and a text model tends to lose fidelity.
Bolt 74B-A9B 24GB is built to run everyday agentic tasks locally on RTX and DGX Spark hardware. It uses native 2-bit quantization to fit in 24GB of memory and run alongside other applications. The "A9B" naming convention in mixture-of-experts models typically denotes a large total parameter count with a smaller number active per token; the name suggests 74 billion total and roughly 9 billion active, which would explain how the model can stay fast despite its size, though that reading is inference from the name rather than something stated. NVIDIA says it balances intelligence and speed against similarly sized models, and that users can point local agent harnesses at it for web research, writing and code debugging without sending workloads to the cloud. It ranks first among open models in its class on PinchBench and second among open models on BrowseComp. It ships on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM, runs on RTX Spark laptops arriving October 16, GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations.
Bolt E4B 4GB targets physical and conversational agents on GeForce RTX, Jetson and RTX/DGX Spark. It pairs natural conversation with multi-turn agentic capability for local voice assistants, interactive characters and embedded interfaces. The 4GB footprint is the striking part: it puts a conversational agent on consumer and embedded hardware, which matters for latency-sensitive voice work where a cloud round trip is the enemy.
The ecosystem story is substantial. NVIDIA says dozens of organizations are post-training Super VL on their own data, tools, policies and workflows. Named examples include Scale AI, which post-trained it for Skydance's vendor credit finance agent and EY's M&A due diligence agent; AfterQuery, for Hebbia's credit agreement analysis; AI21 and Alice, for cybersecurity attack resilience; Mercor, on the APEX-SWE coding benchmark for coding agents; and ModelScope, which enables fine-tuning through its ms-swift framework. Ramp (financial document intelligence), Dream (cybersecurity research and vulnerability analysis), Lila Sciences (scientific discovery and search) and R1 (medical claims analysis and revenue cycle management) are also customizing it. TrainLoop post-trained the model with dermatology data from Nolla, which NVIDIA says just became the first company approved for AI to write initial patient prescriptions.
Platform integrations include Atlassian for agentic search across Jira and Confluence, CrowdStrike for agentic blue teaming in enterprise security, Dassault Systèmes for engineering design and simulation, Perplexity for search and agentic tasks in Perplexity Search and Computer, and Salesforce, which is customizing Nemotron for agents that digitize and automate complex government workflows in secure customer environments through Missionforce Operations. Synopsys design verification software interoperates with the model. The breadth suggests the openness pitch is landing with vendors who want a base they can specialize rather than a closed API.
Openness is the core claim. The model ships with open weights, open datasets and training recipes, and can run wherever security and data-residency rules require. NVIDIA's examples are insurance agents scanning claims for fraud markers and financial services agents checking credit activity against in-house lending criteria, both on premises. Customization runs through NVIDIA NeMo, Python-based pipelines and open datasets.
Training details: Super VL saw about 26 trillion tokens across pretraining and continued pretraining, then post-training on roughly 3.8 million samples plus reinforcement learning across dozens of environments. NVIDIA is releasing open datasets on Hugging Face covering agentic tool use, finance, instruction following, long context, fairness and vision-language understanding, along with reinforcement learning environments in NeMo Gym. The model is available on Hugging Face, OpenRouter, ModelScope and as an NVIDIA NIM microservice with validated configurations and enterprise support refreshed every two weeks with security patches and performance work. Running on NVIDIA Dynamo, NVIDIA says organizations can serve up to 8x more users with 85% lower latency.
Super VL runs across DGX Spark, DGX Station, RTX PRO and RTX systems locally, Jetson at the edge, and data center and cloud environments. The practical reading is a tiered menu: Ultra for the hardest reasoning, Super VL for high-volume document and tool-using agents, the 24GB Bolt for private local work on a single GPU, and the 4GB Bolt for embedded conversation. A likely trade-off for teams is that a 2-bit quantized local model will not match a full-precision data-center deployment on difficult tasks, so agent design should match model tier to workload rather than assume parity.
Why it matters: Teams building agents now have an open, self-hostable option at three hardware tiers, which lowers the barrier for regulated industries that cannot send data to a closed API. Developers evaluating local agent stacks should watch how the Bolt models perform in their own harnesses once E4B lands on October 23. For product teams, the release signals that post-training on proprietary data — not just picking a frontier model — is becoming the main route to specialization, and NVIDIA is courting exactly that workflow.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
NVIDIA today released Nemotron 3.5 Super VL, an open 120-billion-parameter model that delivers the fastest task completion and highest accuracy for specialized always-on agents. Joining it are two new edge models: Nemotron 3.5 Bolt 74B-A9B 24GB, optimized to balance intelligence and speed per GB for local agentic workloads, and Nemotron 3.5 Bolt E4B 4GB, optimized […]