AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop

Collected Oct 15, 2026

Nvidia has released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters aimed at enterprise agents that run continuously in the background, handling text, images, documents and videos. Two smaller Bolt models ship alongside it, targeting local PCs and edge devices. Nvidia detailed the release in a technical blog post and published weights on Hugging Face.

Super VL uses a Mixture-of-Experts architecture in which only 12 billion of the 120 billion parameters are active per token, which keeps inference cost far below what the headline parameter count suggests. Nvidia developed the model with the Nemotron Coalition, a group that includes Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs. The model was trained on more than 25 trillion tokens across pre-training and continued pre-training, with post-training and reinforcement learning adding roughly four million examples across about 100 open training environments.

The model descends from Nemotron 3 Ultra, Nvidia's largest reasoning model. Nvidia says Super VL matches Ultra's accuracy on instruction following, tool calling, multi-step business processes and image analysis using roughly one-fifth the parameters. It also reports that Super VL scores higher than Ultra on agent tests from Hermes and LangChain and beats the earlier Nemotron 3 Super on the coding benchmarks SWE-bench and Terminal-Bench. On one GPU, Nvidia claims up to five times the throughput of comparable open models on agent tasks, citing the Artificial Analysis leaderboard. An independent test by EXO Labs on the DGX Spark found no other model that was both more accurate and faster, according to Nvidia. The Artificial Analysis ranking cited places Super VL well behind leading closed models but ahead of Nemotron 3 Ultra.

Beyond raw architecture, a large part of the pitch is inference efficiency. Nvidia compresses Super VL with its 4-bit NVFP4 format: on Blackwell GPUs the model runs entirely in 4-bit precision, while older cards keep only the weights at 4 bits and do the math at 16 bits, with sensitive layers held at higher precision. Nvidia trained the compressed version to match the uncompressed original's outputs to limit quality loss. A built-in draft component proposes tokens ahead of time that the main model verifies in a single forward pass; on the DGX Spark Nvidia recommends a method called LiLiCorr that evaluates all draft tokens together, which it says speeds text generation by up to 2.3x. Because agents resend the same instructions, tool descriptions and conversation history on every request, Super VL caches those computations instead of rerunning them, and Nvidia's Dynamo inference software routes requests from one session to the GPU already holding that session's cache. On 16 B200 GPUs, Nvidia says this raises concurrent users eightfold and cuts time to first token by 85 percent.

Vision work is aimed at everyday business material: OCR, tables, charts, multi-page documents and video. Super VL can read a chart, connect it to a table elsewhere in the same document and summarize both without a separate vision model, Nvidia says. On the chart benchmark CharXiv it scores 80.8 percent, up from 63 percent for the older Nemotron 3 Nano Omni, and that rises to 84 percent when the model can also run Python code. Nvidia says it ranks among the best open models on the OCRBench v2 leaderboard.

The two Bolt variants are the lightweight arm of the launch. Bolt 74B-A9B 24GB is a hybrid Mixture-of-Experts model that activates 9 of its 74 billion parameters per step and is squeezed into 24 GB of memory using aggressive 2-bit weight quantization. Nvidia positions it for local agents that take multiple steps, call tools or search the web, and for tasks such as writing and debugging code or working out which PC components are compatible. It is optimized for the RTX Spark chip for AI-capable Windows laptops and the DGX Spark workstation with up to 128 GB of shared memory, and also runs on GeForce RTX PCs, DGX Station and RTX PRO workstations, supporting llama.cpp, Ollama, LM Studio and vLLM. Bolt E4B 4GB, planned for later this month, targets voice assistants, interactive characters and device interfaces on GeForce RTX cards, Jetson modules and RTX or DGX Spark hardware.

The 24 GB requirement is a significant practical constraint. Memory at that level comes mainly with high-end, expensive hardware, so the local agent is realistically aimed at developers and owners of powerful workstations rather than mainstream PCs, at least at current GPU prices. A likely trade-off of 2-bit quantization is that the smallest local variant will trail the full model on harder reasoning, even if it handles routine agent loops well.

Early adopters, which Nvidia says number in the dozens and cluster in document-heavy, data-sensitive industries, include Scale AI, which adapted the model for an agent handling vendor credits at media company Skydance, with costs expected to fall to one-fortieth and response time to one-fifth; Ramp, which turns photos of receipts into structured data and expects ten times faster responses; EY for acquisition due diligence; and Hebbia for credit agreement analysis. In cybersecurity, AI21 fine-tuned Super VL for security firm Alice, which uses it to analyze threats and vulnerabilities, with Nvidia reporting a 40 percent accuracy improvement, while Dream hunts for vulnerabilities and CrowdStrike is building defensive agents. Salesforce, R1, Perplexity, Dassault Systèmes and Mercor are also named as users.

Nvidia is releasing the weights in base, post-trained and 4-bit compressed versions, plus datasets, training recipes and reinforcement learning environments, which is more than most open-weight releases provide. Several datasets go up on Hugging Face covering tool use, instruction following, long context and image understanding, along with finance, safety and fairness sets. One covers languages spoken in India, built with the Indian Institute of Technology Madras and its AI4Bharat lab, containing about 2.83 trillion tokens across 22 Indian languages; fine-tuned on it, Super VL beats larger specialized models such as Sarvam-105B on Indian language benchmarks, according to Nvidia. Weights are listed for Nemotron 3.5 Super VL 120B-A12B and Nemotron 3.5 Bolt 74B-A9B 24GB. Through Nvidia NIM, enterprises get the model as a deployable microservice with validated configurations and support, with updates planned every two weeks.

The open release is paired with tight optimization for Nvidia's own stack, spanning DGX Spark, DGX Station, RTX PRO, RTX, Jetson modules, data centers and the cloud. In practice that means the cheapest path to the advertised performance is Nvidia hardware. This is a familiar pattern: permissive-enough weights grow the installed base of model users, and the compute bill follows, even as rivals such as Meta, Google, Anthropic and OpenAI build custom silicon and tune models for it. Nvidia recently invested $12.9 billion in Hugging Face, signaling that it wants the open-model ecosystem itself to keep expanding. An open question is whether the ecosystem will keep porting these models to non-Nvidia accelerators fast enough to blunt that pull, and how much of Super VL's efficiency depends on Blackwell-specific formats and Dynamo rather than the model weights alone.

Why it matters: Teams that need agents to comb through documents or images under their own terms of service now have an open multimodal option they can fine-tune on proprietary data and self-host, with published recipes lowering the cost of adapting it. The catch is hardware: the 120B model runs best on Blackwell-class GPUs, and even the local Bolt variant needs 24 GB of memory, so budget for workstations or rented compute. Expect the strongest returns for document-heavy finance, insurance and security workflows, where Nvidia's cited early adopters are already fine-tuning.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

With Nemotron 3.5 Super VL, Nvidia is releasing an open multimodal model with 120 billion parameters for enterprise agents. Two compact Bolt models bring agents to local machines. Dozens of organizations are already fine-tuning Super VL, but getting the most out of it means buying Nvidia hardware. The article Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop appeared first on The Decoder .