Skip to content
Neural Daily

ModelsNews

Nemotron 3.5 Super VL: NVIDIA's open model for agents

NVIDIA released Nemotron 3.5 Super VL on October 15: an open 120-billion-parameter multimodal model that activates 12 billion parameters per token. Two smaller Bolt models arrive alongside it to run agents on local hardware.

By: Mikola Máté Patrik · · 2 min read

NVIDIA Nemotron 3.5 Super VL announcement graphic
Image: NVIDIA

What shipped

Nemotron 3.5 Super VL is a mixture-of-experts (MoE) model with 120 billion total and about 12 billion active parameters. It reasons over text, images, documents and video while calling tools, and NVIDIA pitches it at always-on, specialised agents (NVIDIA).

NVIDIA calls it the fastest open agentic model in its class, five times faster than the next best. The technical blog lists up to 1M tokens of context and native MTP speculative decoding (NVIDIA Technical Blog).

Chart of Artificial Analysis Intelligence Index versus output speed for open-weight models, with Nemotron 3.5 Super VL far right
NVIDIA / Artificial Analysis

Source: NVIDIA / Artificial Analysis. Intelligence Index versus output speed for open-weight models at 100k input tokens.

Bar chart of Artificial Analysis Intelligence Index scores including Nemotron 3.5 Super VL at 26
NVIDIA / Artificial Analysis

Source: NVIDIA / Artificial Analysis. Artificial Analysis Intelligence Index v4.3.2 scores; Super VL scores 26.

What it is for

Its vision skills target what agents actually meet: OCR, tables, charts, PDFs and web pages. Weights, datasets and training recipes are open, so teams can post-train it on their own data and run it on their own infrastructure.

NVIDIA says it was trained on about 26 trillion tokens, then post-trained with roughly 3.8 million samples and RL across dozens of environments. Named adopters include Atlassian, CrowdStrike, Perplexity and Salesforce.

Bolt: agents on your own machine

Nemotron 3.5 Bolt 74B-A9B 24GB uses native 2-bit quantisation to fit in 24 GB of memory on RTX and DGX Spark systems. The smaller Bolt E4B 4GB targets voice and embedded agents and is due on October 23.

Availability

Super VL is available now on Hugging Face, OpenRouter, ModelScope and as an NVIDIA NIM. Bolt 74B supports llama.cpp, Ollama, LM Studio, Unsloth and vLLM.

Key takeaways

  • Open multimodal agent model with 120B total and 12B active parameters.
  • NVIDIA claims a 5x speed lead in its class; it scores 26 on the AA index.
  • Bolt models target local hardware; E4B ships October 23.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles