Today for AI
HOT RADAR
llmHEAT 119.8°

AI CURATED HOT STORY · 10/15/2026

NVIDIA Open-Sources Nemotron 3.5 Super VL: 120B MoE Model Doubles Intelligence with Near-Lossless NVFP4 Quantization

7 reports archived7 independent sourcesupdated 10/16/2026, 00:09:44
Synthesis & Latest Updates
7 Sources Cross-Validated

NVIDIA has open-sourced Nemotron 3.5 Super VL, a 120B-parameter (12B active) Mixture-of-Experts multimodal model that tops the US open-weight leaderboard with an intelligence score of 26.5. It doubles the intelligence of its predecessor while reducing output tokens by approximately 60%, significantly enhancing inference efficiency.

Designed for persistent enterprise agents, the model supports text, image, document, and video understanding with a 1M context window. Its NVFP4 quantized version maintains near-lossless performance (within 0.1 points of BF16). NVIDIA also released two lightweight Bolt models (74B-A9B and E4B) optimized for local workstations and edge devices, respectively, demonstrating high parameter efficiency.

This release completes NVIDIA's full-stack open model matrix from data center to edge. With Day 0 support in vLLM and customization via the NeMo framework, developers can leverage up to 5x higher throughput on single GPUs compared to similar open models, facilitating the deployment of high-accuracy visual reasoning and tool-calling workflows.

LATEST/Nemotron 3.5 Super VL is now live on OpenRouter with 1M context support. vLLM offers Day 0 compatibility, enabling immediate deployment and customization via the NeMo framework for enterprise agent workflows.

HEAT TRENDHourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 7 reports · Latest first
  1. The DecoderT2·85 pts
    • Nemotron 3.5 Super VL uses a MoE architecture with 120B total parameters but activates only 12B per token, balancing performance and efficiency.
    • Designed specifically for 'always-on' enterprise agents, it natively supports multimodal reasoning across text, images, documents, and video.
    • Companion Bolt 74B-A9B and Bolt E4B models target local workstations (24GB VRAM) and edge devices (4GB VRAM), respectively.
  2. Artificial Analysis (@ArtificialAnlys)T1·86 pts
    • Nemotron 3.5 Super VL leads US open-weight models with an Intelligence Index score of 26.5, surpassing competitors like Thinking Machines' Inkling Small.
    • The new model doubles the intelligence of its predecessor while reducing output tokens per task by approximately 60%, enhancing inference efficiency.
    • NVFP4 quantized weights show near-lossless performance, scoring within 0.1 points of BF16, validating low-precision formats for maintaining capability.
    • The newly launched Bolt 74B-A9B model achieves a score of 21 with only one-seventh of the parameters of the flagship Ultra, offering high cost-efficiency.
  3. OpenRouter (@OpenRouter)T1·68 pts
    • NVIDIA Nemotron 3.5 Super VL is an open-source 120B-A12B multimodal model now available on OpenRouter.
    • The model supports a 1 million token context window, optimized for long-running agent scenarios requiring complex reasoning.
    • It features cross-modal capabilities handling text, images, documents, GUI screenshots, and video, suitable for automation tasks.
  4. IT之家 · 智能时代T2·84 pts
    • Nemotron 3.5 Super VL uses an MoE architecture with 120B total but only 12B active parameters, balancing capability and efficiency.
    • Optimized for multi-step AI agent tasks, achieving up to 5x higher throughput than comparable open models on a single GPU.
    • Integrates text, image, document, and video understanding, supporting enterprise-level post-training customization via the NeMo toolkit.
  5. NVIDIA BlogT1·78 pts
    • Nemotron 3.5 Super VL is a 120B-parameter open model optimized for highest accuracy and fastest task completion in always-on agents.
    • Two new edge models, Bolt 74B-A9B (24GB) and Bolt E4B (4GB), are introduced to balance intelligence and speed for local workloads and conversational agents.
    • The update completes the Nemotron family, joining Lightning (throughput) and Ultra (reasoning) to cover the full range of agent deployment scenarios.
  6. vLLM 官方博客T1·68 pts
    • Nemotron 3.5 Super VL is an open-source multimodal model designed for long-running agents, unifying visual understanding and tool use.
    • Utilizes a 120B total parameter MoE architecture with 12B active parameters to balance accuracy and inference efficiency.
    • Supports custom post-training using the NVIDIA NeMo framework with private data, tools, and policies.
    • vLLM offers Day 0 support, allowing developers to immediately deploy and test via official cookbooks.
  7. Artificial AnalysisT1·86 pts
    • Nemotron 3.5 Super VL tops the US open-weight model rankings with a score of 26.5, outperforming Inkling Small and Muse Glimmer.
    • The new model doubles the intelligence of its predecessor while reducing output tokens per task by approximately 60%.
    • It supports NVFP4 quantization with minimal performance loss (<0.1 points compared to BF16), facilitating efficient deployment.
    • The companion Bolt 74B-A9B achieves near-flagship performance using only ~1/7th of total parameters and ~1/6th of active parameters.