Breaking

NVIDIA Open-Sources Nemotron 3.5 Super VL as 120B Partner-Trained Agent Model

October 16, 2026 at 18:04 EDT

Foundation Models
AI Agents
Open Source

NVIDIA has released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model designed for enterprise agent deployment, with over 100 partners already post-training specialized variants for finance, security, healthcare, and software development. The model, launched October 15, 2026, combines text, image, document, and video understanding with agentic reasoning capabilities in a single system, achieving up to 5x higher throughput than comparable open models on a single GPU.

NVIDIA · October 2026

120B Open Nemotron Model Sees Rapid Partner Specialization

NVIDIA released Nemotron 3.5 Super VL as a fully open multimodal agent model. Over 100 partners have already post-trained specialized versions, delivering 8% accuracy gains and 33x cost reductions in early deployments.

5×

Higher throughput vs. open models

100+

Partners customizing the model

33×

Cost reduction (Deel payroll agent)

40×

Lower inference cost (Skydance finance)

Inference Throughput: Nemotron 3.5 Super VL vs. Other Open Models

Other Open
Models

1×

Nemotron
3.5 Super VL

5×

Deel Payroll QA

+8%

Accuracy vs. Claude Opus 5.5 on 8-country benchmark

Skydance Vendor Credit

40× Lower

Inference cost via Scale AI post-training

Early Partner Applications

Finance: Skydance (vendor credit), EY (M&A due diligence), Ramp (receipt extraction), R1 (medical claims)

Security: Alice (threat detection), CrowdStrike (blue teaming), Dream (vulnerability analysis)

Development: Mercor (coding agents), Atlassian (agentic search), Synopsys (design verification)

Other: Perplexity (search), Salesforce (workflow), Dassault Systèmes (engineering), Lila Sciences (discovery)

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.