中文 | English
← Back to article list

NVIDIA Unveils Nemotron 3.5 Super VL and Bolt: 120B Multimodal Agentic Reasoning, Enterprise Document Intelligence, and On-Device Local Inference

LLM & Multimodal

Introduction: The Maturation of Enterprise Multimodal Agentic Reasoning

As enterprise multi-agent workflows transition from experimental proofs-of-concept into mission-critical infrastructure, traditional text-centric Large Language Models (LLMs) confront clear operational ceilings. Enterprise data is inherently non-linear and visual: dense multi-page annual reports, intricate architectural schematics, tabular financial spreadsheets, and complex software graphical user interfaces (GUIs). Routing these documents through crude Optical Character Recognition (OCR) parsers strips away geometric orientation and relational hierarchy, resulting in catastrophic loss of context. On October 15, 2026, NVIDIA addressed this computational bottleneck by unveiling its next-generation open multimodal model portfolio, headlined by the flagship Nemotron 3.5 Super VL (120B) alongside the edge-optimized Nemotron 3.5 Bolt (74B-A9B) series, establishing a comprehensive standard for document intelligence, agentic tool execution, and secure on-premise deployment.

Core Technical Architecture: High-Throughput Inference and Specialized Post-Training

Engineered for enterprise scale, the Nemotron 3.5 family demonstrates notable advancements in architecture, processing velocity, and post-training adaptability:

  • Unified Multimodal Agentic Foundation (Super VL 120B): Super VL integrates language, tabular layout, and high-resolution visual comprehension within a monolithic 120-billion-parameter architecture. Pre-trained on roughly 26 trillion tokens across NVIDIA Blackwell infrastructure and augmented with 3.8 million domain post-training sequences, Super VL directly interrogates multi-page PDFs, structured balance sheets, and interactive web environments without delegating visual understanding to auxiliary subsystems.
  • Unprecedented Output Velocity (625 Tokens/s): Independent evaluations conducted by Artificial Analysis indicate that when deployed on an NVIDIA Blackwell B200 four-GPU configuration, Super VL delivers sustained generation throughput of 625 output tokens per second under a 100,000-token input load. This fivefold speedup over comparable open multimodal models eliminates conversational latency during intricate multi-step agent reasoning loops.
  • Dramatic Task Specialization via NeMo Post-Training: Early deployment data from partners including Scale AI, EY, and Hebbia highlights the model's adaptability. When specialized on institutional data, Super VL achieved a remarkable jump in medical insurance adjudication accuracy from 28% to 72%, substantially outpacing Claude Opus 5 (38%). Similarly, security attack resilience climbed from 35% to 76%, demonstrating the practical value of open-weight fine-tuning over generic public APIs.
  • Local Edge Agent Pioneer: Nemotron 3.5 Bolt: Designed for local coding, research, and sensitive workloads, Bolt 74B-A9B features a 74-billion-parameter Mixture-of-Experts (MoE) design that activates only 9 billion parameters per token. With native 2-bit quantization, Bolt fits comfortably into a standard 24GB VRAM footprint on consumer workstations, enabling sophisticated offline agentic execution with zero data egress.

Strategic Value: Data Residency, Commercial Freedom, and Governance

Released under NVIDIA's flexible OpenMDW license, Nemotron 3.5 allows unrestricted enterprise fine-tuning, commercial self-hosting, and weight redistribution:

  • Rigorous Legal and Financial Due Diligence: Financial institutions and corporate legal teams can deploy customized Super VL agents to perform line-by-line contract analysis and transaction audits completely within their private virtual clouds, adhering strictly to global data sovereignty mandates without compromising performance.
  • Frictionless Local Developer Assistance: With Bolt models running directly on engineering workstations, software organizations can provide engineers with continuous coding assistants and automated testing tools that inspect internal repositories with absolute operational privacy.

Integration with FD Studio: Powering the Intelligent Creative Studio

As modern creative workflows evolve toward agent-orchestrated multimodal production, FD Studio has integrated NVIDIA Nemotron 3.5 into its all-in-one generative suite:

  • Omnimodal Director Agent for Production Orchestration: FD Studio utilizes Nemotron 3.5 Super VL as an intelligent creative director. When creators submit preliminary pitch decks featuring concept art, narrative prose, and sonic moods, the Super VL agent breaks down the narrative into coherent cinematic storyboards, framing directions, and synchronized shot sequences.
  • Hybrid Edge-Cloud Creative Processing: By bridging cloud-based Super VL models for macro narrative planning with local Bolt instances for responsive prompting and parameter fine-tuning, FD Studio delivers an adaptive, high-throughput workspace that empowers creative studios to produce next-generation digital content with unmatched speed and creative control.