NVIDIA Releases Nemotron 3.5 Super VL, an Open-Weight 120B Model Built for Always-On AI Agents

NVIDIA has released a new family of open-weight AI models designed to power AI agents that run nonstop in the background. The headline model, Nemotron 3.5 Super VL, packs 120 billion parameters into a mixture-of-experts design where only 12 billion are active per token, joined by two compact “Bolt” models built for local PCs and edge devices. The company announced the launch on October 15, 2026, publishing the weights on Hugging Face the same day.

Aimed at always-on enterprise agents

Super VL is built for specialized agents that run continuously, sometimes for days or weeks. Typical jobs include coding, web research, document processing, and cybersecurity triage.

Vision is a particular focus. The model’s vision targets the materials agents meet at work: OCR, tables, charts, multi-page documents, and video. It accepts JPEG or PNG images and MP4 videos up to two minutes long, and responds in text. Beyond English, NVIDIA focused post-training on French, German, Italian, Japanese, Spanish, and Chinese.

The key numbers, and whose they are

According to NVIDIA, Super VL was distilled from Nemotron 3 Ultra, the company’s largest reasoning model, and matches Ultra’s accuracy on agent tasks such as instruction following, tool calling, multi-step business workflows, and image analysis with roughly one-fifth the parameters. In agent tests from Hermes and LangChain, NVIDIA says it actually scores higher than Ultra, and it beats its predecessor Nemotron 3 Super on the coding benchmarks SWE-bench and Terminal-Bench.

On speed, NVIDIA claims up to five times the throughput of other open models of similar size on agent tasks, citing the Artificial Analysis leaderboard, with Super VL topping 600 output tokens per second on Baseten’s endpoint. An independent test by EXO Labs on the DGX Spark found that no other tested model was both more accurate and faster than Super VL.

On visual benchmarks, NVIDIA reports 80.8 percent on the chart benchmark CharXiv, rising to 84 percent when the model can also run Python code, compared with 63 percent for the older Nemotron 3 Nano Omni. The model card lists 52.3 on OSWorld-Verified and 68.34 on OCRBench v2 in English.

Two caveats matter. Most of the numbers above come from NVIDIA or its partners, even when the company points to the Artificial Analysis leaderboard. And as of the evening of October 15, there were no independent benchmarks of the Bolt models yet.

Two models now, a third on October 23

Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB are available now; Nemotron 3.5 Bolt E4B 4GB is scheduled for October 23. The specs, per NVIDIA’s announcement:

• Super VL 120B-A12B: 120 billion total parameters, 12 billion active per token, 1 million token context window. Deployment options include a single DGX Spark, one H200 GPU, two H100s, or one GB200.
• Bolt 74B-A9B 24GB: 74 billion total, 9 billion active, 256,000 token context. A hybrid mixture-of-experts design with weights compressed to about 2 bits per parameter, so the model fits in 24 GB of memory. Tested on a PC with a GeForce RTX 5090 and on DGX Spark.
• Bolt E4B 4GB: details undisclosed so far, aimed at voice assistants and device interfaces on GeForce RTX cards, Jetson modules, and RTX/DGX Spark systems.

The weights come in a base version, a post-trained version, and a 4-bit compressed version using NVIDIA’s NVFP4 format. Unusually for an open-weight release, NVIDIA also published training datasets, recipes, and reinforcement-learning environments: more than 25 trillion training tokens, roughly four million post-training examples, plus open datasets on Hugging Face.

The license is OpenMDW-1.1, which allows commercial use, modification, and redistribution as long as the license and original notices are included.

Early adopters and the business angle

NVIDIA says dozens of organizations are already fine-tuning Super VL, mainly in industries with heavy document loads and sensitive data. Scale AI adapted it for vendor credit processing at media company Skydance, which NVIDIA says will cut inference cost to one-fortieth and latency to one-fifth. Ramp uses it to turn receipt photos into structured data and expects up to 50 percent higher throughput. AI21 fine-tuned it for security firm Alice’s threat intelligence and vulnerability analysis, which NVIDIA says improved accuracy by 40 percent. Named adopters also include EY, CrowdStrike, Salesforce, Dassault Systèmes, and Perplexity.

The strategy is straightforward: open models that companies can fine-tune on proprietary data and run on their own servers keep sensitive documents in-house, and they drive demand for the hardware they run on. NVIDIA developed Super VL with the Nemotron Coalition, a group of AI labs that includes Mistral AI and Perplexity.

Why it matters

This release continues the shift from the parameter-count race to efficiency. Super VL is designed to match a much larger model’s accuracy at a fraction of the compute, while Bolt squeezes a 74-billion-parameter agent model into a single workstation’s memory. If the speed claims hold up under independent testing, open-weight models get meaningfully closer to being practical for long-running business agents that enterprises host themselves. The missing piece, as always with vendor-launched models, is independent verification: treat NVIDIA’s benchmark leads as a claim to check, not a verdict, and the same goes for the Bolt models, which had no third-party results at launch.

Sources

• The Decoder, “Nvidia’s Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop,” October 15, 2026: https://the-decoder.com/nvidias-nemotron-3-5-super-vl-is-an-open-model-built-to-run-enterprise-ai-agents-nonstop/
• Maccus, “NVIDIA squeezes a 74B-parameter open model into 24 GB and releases a 120B agent model that reads PDFs,” October 15, 2026: https://maccus.fr/en/nvidia-nemotron-3-5-open-models/