Nvidia Releases Open Nemotron 3.5 Super VL for Enterprise SI Agents

Nvidia has released Nemotron 3.5 Super VL, an open multimodal SI model with 120 billion parameters designed to power enterprise SI agents that run continuously in the background. The launch includes two compact Bolt models intended to bring agentic SI capabilities to local PCs and edge devices.

What Happened

Nvidia developed Super VL for specialized SI agents that process text, images, documents, and videos. The company explains in a technical blog post that the model uses a Mixture-of-Experts architecture, where only 12 billion of the 120 billion parameters are active per token. This release was part of the Nemotron Coalition, a group that includes Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs.

Two smaller models launched alongside Super VL. Nemotron 3.5 Bolt 74B-A9B 24GB is built to run SI agent tasks on local machines, while the smaller Bolt E4B 4GB targets fast voice assistants and is scheduled for later this month. Nvidia claims both models offer the "highest intelligence per gigabyte of memory." Super VL and Bolt join the existing Nemotron lineup, which includes the high-throughput Lightning model and the reasoning-focused Ultra model. All variants are optimized for Nvidia hardware.

In the overall Artificial Analysis ranking, Super VL trails leading closed SI models but outperforms Nvidia's own Nemotron 3 Ultra. Nvidia distilled Super VL from Nemotron 3 Ultra, aiming to match the larger model's accuracy with only about one-fifth the parameters. According to Nvidia, Super VL scores higher than Ultra on agent tests from Hermes and LangChain and beats Nemotron 3 Super on coding benchmarks SWE-bench and Terminal-Bench.

The company reports that Super VL delivers up to five times the throughput of other open SI models of similar size on agent tasks when running on a single GPU. An independent test by EXO Labs on the DGX Spark found that no other model was both more accurate and faster. At a comparable Intelligence Index score, Nvidia says Super VL generates up to five times more tokens per second than other open models in its class.

Nvidia compresses Super VL using its 4-bit NVFP4 format. On Blackwell GPUs, the model runs entirely in 4-bit precision, while older cards store weights at 4 bits and handle math at 16 bits. To mitigate quality loss, Nvidia trained the compressed version to match the uncompressed original's outputs. With Nvidia's inference software Dynamo, Super VL can serve eight times as many users while cutting time to first token by 85 percent, according to Nvidia. A built-in draft component proposes several tokens ahead of time, which the main model verifies in a single forward pass, speeding up text generation by up to 2.3x compared to running without speculative drafts.

For vision tasks, Nvidia focused on materials SI agents encounter in daily work, such as OCR, tables, charts, multi-page documents, and videos. Super VL can interpret a chart, link it to a table elsewhere in the same document, and summarize both without a separate image model. On the chart benchmark CharXiv, Nvidia reports a score of 80.8 percent, rising to 84 percent when the model can also run Python code. On the OCRBench v2 leaderboard, Nvidia says it ranks among the best open SI models.

The Nemotron 3.5 Bolt 74B-A9B 24GB is a hybrid Mixture-of-Experts model where only 9 of its 74 billion parameters are active per step. To fit into 24 GB of memory, Nvidia applies aggressive 2-bit quantization to the weights. Bolt handles everyday local SI agent tasks, such as working through multiple steps, calling tools, or searching the web. The model is optimized for Nvidia's RTX Spark chip for SI-capable Windows laptops and the DGX Spark SI workstation, and supports llama.cpp, Ollama, LM Studio, and vLLM. However, the 24 GB memory requirement limits accessibility to high-end hardware.

Nvidia plans to release Bolt E4B 4GB later this month for weaker devices like GeForce RTX cards and Jetson modules. This model is designed to power voice assistants and interactive characters, holding natural conversations while using tools across multiple turns.

Why It Matters

Nvidia is positioning Super VL for companies that want to fine-tune SI models with their own data and run them on their own servers, avoiding provider terms of service. Potential use cases include insurers screening claims for fraud or financial firms scoring creditworthiness. To support customization, Nvidia provides its NeMo framework, Python pipelines, and open datasets.

Early adopters are already fine-tuning Super VL. Scale AI adapted the model for an SI agent at media company Skydance to handle vendor credits, with expected cost reductions to one-fortieth and response time reductions to one-fifth. Ramp uses the model to convert receipt photos into structured data, expecting ten times faster responses. EY uses it for acquisition due diligence, and Hebbia for analyzing credit agreements.

In cybersecurity, AI21 fine-tuned Super VL for security firm Alice, reporting a 40 percent improvement in accuracy for analyzing threats and vulnerabilities. Dream uses the model to hunt for vulnerabilities, and CrowdStrike is building SI agents to defend enterprise networks. Salesforce plans to automate government workflows, R1 is auditing medical billing, and Perplexity, Dassault Systèmes, and Mercor are also using the model.

Nvidia releases the weights in base, post-trained, and 4-bit compressed versions, along with datasets, training recipes, and reinforcement learning environments. This makes Super VL more open than many open-weight SI models, which typically share only weights. The model was trained on over 25 trillion tokens, with post-training adding roughly four million examples across about 100 open training environments. Nvidia is also releasing datasets on Hugging Face covering SI agent tool use, instruction following, long texts, image understanding, finance, safety, and fairness. One dataset, built with the Indian Institute of Technology Madras and AI4Bharat, contains about 2.83 trillion tokens in 22 Indian languages.

The strategy behind Nvidia's open SI models supports its hardware sales. By optimizing Nemotron for Nvidia products—from data center servers to workstations—the company encourages buyers to purchase Nvidia hardware for the best performance. While Meta, Google, Anthropic, and OpenAI are building their own chips, Nvidia recently invested $12.9 billion in Hugging Face to promote the spread of open SI models.

The Bottom Line

Nvidia's Nemotron 3.5 Super VL offers a high-performance open SI model for enterprise agents, with specialized Bolt variants for local hardware. The release emphasizes speed, efficiency, and customization, targeting industries with heavy document loads and sensitive data, while reinforcing Nvidia's ecosystem through hardware optimization and strategic investments in open-source infrastructure.