Nvidia Debuts Nemotron 3.5 Super VL, an Open Multimodal Model for Enterprise AI Agents
Nvidia has launched Nemotron 3.5 Super VL, a 120B open multimodal model engineered for continuous enterprise agent workflows, alongside local Bolt models for PC and edge hardware.
Nvidia has unveiled Nemotron 3.5 Super VL, a new open multimodal AI model packing 120 billion parameters, designed specifically to power persistent enterprise agents. Created alongside the Nemotron Coalition—which brings together partners such as Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs—the system is geared toward tasks that run continuously, handling inputs across text, charts, images, documents, and video.
The model relies on a Mixture-of-Experts architecture that activates only 12 billion parameters per token. In Artificial Analysis rankings, Super VL trails the top closed frontier options by a wide margin, but Nvidia reports that it outperforms its own Nemotron 3 Ultra. To achieve this balance, Nvidia distilled Super VL from Nemotron 3 Ultra, matching Ultra's accuracy on multi-step workflows, tool calls, instruction following, and visual understanding with roughly one-fifth of the parameters. In benchmarks from Hermes and LangChain, Super VL scored higher than Ultra. While it proved slightly less accurate than rival models on coding benchmarks like SWE-bench and Terminal-Bench, Nvidia reports it completed those coding evaluations four times faster.
High Throughput Through Compression and Caching
To make non-stop background operation cost-effective, Nvidia applied extensive optimization techniques. Nemotron 3.5 Super VL uses a 4-bit NVFP4 format, enabling Blackwell GPUs to run calculations entirely in 4-bit precision, while older architectures store weights at 4 bits and compute at 16 bits. High-precision representation is preserved for sensitive model components.
Inference speed also benefits from speculative drafting. A built-in draft component anticipates future tokens that the primary network validates in a single pass, while the alternative LiLiCorr approach evaluates tokens collectively on DGX Spark systems to yield up to 2.3 times faster text output. Because autonomous agents frequently repeat system prompts, tool schemas, and chat records, Super VL avoids redundant work through prefix caching. Paired with Nvidia's Dynamo inference engine, requests are routed to GPUs that already store the corresponding session cache. Running on 16 B200 GPUs, this setup cuts time to first token by 85 percent and supports eight times more concurrent users.
Super VL handles document understanding directly without relying on an external vision network. On the CharXiv chart benchmark, the model scored 80.8 percent—rising to 84 percent when paired with Python code execution—compared to 63 percent for the earlier Nemotron 3 Nano Omni. It also ranks near the top of the OCRBench v2 leaderboard among open models.
Local Execution with Nemotron Bolt Models
Alongside the flagship model, Nvidia introduced two edge-oriented variations. The Nemotron 3.5 Bolt 74B-A9B 24GB is a compact hybrid Mixture-of-Experts release activating 9 billion out of 74 billion parameters per step. Using aggressive 2-bit quantization, the model fits into 24 GB of memory to run tool calling, web search, coding, and debugging directly on workstations and PCs. It supports engines like llama.cpp, Ollama, LM Studio, and vLLM, and is tailored for setups like the DGX Spark, RTX PRO, GeForce RTX systems, and the RTX Spark chip for Windows laptops.
A smaller companion, the Bolt E4B 4GB, is scheduled to arrive later this month. Aimed at GeForce RTX cards, Jetson edge modules, and RTX/DGX Spark hardware, it targets low-latency voice agents and contextual app assistants. However, Nvidia's 24 GB local agent remains bound to high-end hardware due to current GPU costs.
Super VL has already seen deployment across enterprise sectors. Scale AI implemented it for Skydance to process vendor credits, aiming for response times one-fifth of prior setups and costs down to one-fortieth. Ramp uses it to extract data from receipt images, while EY and Hebbia apply it to due diligence and credit agreement reviews. In cybersecurity, AI21 reported a 40 percent accuracy boost applying the model for Alice, with Dream, CrowdStrike, Salesforce, and R1 also integrating it.
What it means for developers
For engineering teams, Super VL offers an accessible blueprint for self-hosted agent infrastructure. Nvidia has released the base, post-trained, and 4-bit compressed weights on Hugging Face, alongside training recipes and reinforcement learning environments. Pre-training covered over 25 trillion tokens, followed by four million post-training examples across approximately 100 environments. Nvidia also published specialized datasets, including an Indic language collection built with IIT Madras and AI4Bharat comprising 2.83 trillion tokens across 22 languages, on which fine-tuned Super VL models outperformed Sarvam-105B.
Teams can deploy the model on premises or in the cloud via Nvidia NIM microservices to maintain strict data ownership without relying on third-party terms of service. For teams that want to benchmark such open-weight architectures against hosted industry alternatives, developers can try top AI models cheaply through one API at https://apixoai.online.
Ultimately, while open releases empower developers to run agents without provider restrictions, they also reinforce the computing demands tethered to Nvidia hardware. With competing platform giants building proprietary silicon, releasing models heavily tuned to Blackwell and RTX systems helps Nvidia secure hardware demand across enterprise and workstation deployments.
Source: Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
