What Is Nemotron 3.5 Super VL? NVIDIA's Open Agent Model
The short answer
Nemotron 3.5 Super VL is NVIDIA’s open 120B (12B active) multimodal model for enterprise agents, released October 15, 2026, alongside two edge models, Bolt 24GB and Bolt 4GB. Its pitch is efficiency rather than topping leaderboards: NVIDIA claims up to 5x the throughput of other open models in its size class on a single GPU, at accuracy close to its much larger Nemotron 3 Ultra. It reads documents, charts, web pages and video as well as text, and it is cheap to call — $0.15 / $0.50 per million tokens on Together AI.
Key facts
| Nemotron 3.5 Super VL | Bolt 74B-A9B 24GB | Bolt E4B 4GB | |
|---|---|---|---|
| Released | Oct 15, 2026 | Oct 15, 2026 | Due Oct 23, 2026 |
| Parameters | 120B total / 12B active | 74B / 9B active (compressed from Super VL) | ~4B effective |
| Architecture | Mamba-2–Transformer hybrid MoE, multi-token prediction | Same family, native 2-bit | — |
| Context | Up to 1M tokens | Up to 256K tokens | — |
| Input | Text, image, document, video | Text, image, video | Text/voice |
| Runs on | Data centre GPUs, DGX Spark/Station, RTX PRO | RTX/DGX Spark, GeForce RTX (24 GB) | GeForce RTX, Jetson, Spark |
| Licence | OpenMDW 1.1 (commercial use allowed) | OpenMDW 1.1 | — |
| API price | $0.15 in / $0.02 cached / $0.50 out (Together) | local | local |
What it is good at
NVIDIA targets five workloads: coding (test, debug, fix), web research, document intelligence, cybersecurity triage and video understanding. The visual training focuses on what agents actually meet — OCR, tables, charts and multipage PDFs. NVIDIA reports 80.8% on CharXiv chart reasoning (Nemotron 3 Nano Omni: 63%) and a place among the top open models on OCRBench v2.
Early customers post-train it for narrow jobs: Scale AI for vendor credit processing at Skydance (NVIDIA cites 40x lower inference cost and 5x lower latency), Ramp for turning receipt images into JSON, AI21 and Alice for threat intelligence (+41% accuracy). Atlassian, CrowdStrike, Perplexity and Salesforce are integrating it. These are vendor-reported results.
Why it is fast
- 12B active parameters out of 120B keep per-token compute low.
- Multi-token prediction gives it a built-in drafter, and NVIDIA’s open LiLiCorr speculative decoding adds up to 2.3x faster generation on DGX Spark.
- NVFP4 weights with quantization-aware distillation keep accuracy at 4-bit on Blackwell GPUs.
- Context reuse across agent turns avoids re-processing the same prompt on every tool call.
Should you use it?
- Yes for high-volume document and multimodal agents where cost per task matters more than peak intelligence — at $0.50 per million output tokens it undercuts most closed models by 10x or more.
- Yes if you need to fine-tune on private data and deploy on your own GPUs; weights, data and recipes are all open.
- Look elsewhere for frontier coding or hard reasoning: closed flagships and the largest open models still lead. Compare current API prices and the best open-weight models ranked by value.
- For a laptop or desktop agent, Bolt 24GB is the one to try: it fits in 24 GB alongside other apps and runs in Ollama and LM Studio.
For the earlier tier, see what Nemotron 3.5 Lightning is; for the other big open release of the same day, Meta’s Muse Spark Open.
Last verified: October 16, 2026.