NVIDIA launches high-speed Nemotron models for agentic AI

NVIDIA's Nemotron 3.5 family is a lineup of open models built specifically for autonomous agents. The flagship is Super VL, a 120B multimodal model designed for high-speed enterprise automation, accompanied by Bolt 74B-A9B, an edge-ready model that runs locally on RTX workstations. Instead of chasing benchmark records on extreme reasoning, these models prioritize raw throughput and practical task execution.

What it does and how it works

Super VL combines vision, document understanding, and tool use into a single system. In typical agent setups, developers string together separate OCR or vision models with text models, which creates lag and context loss. Super VL natively processes complex PDFs, spreadsheets, charts, and video files, extracting insights and triggering tool calls without bouncing data across different models.

Under the hood, Super VL uses a 120B mixture-of-experts (MoE) architecture that only activates 12B parameters per token. That architectural choice is how it delivers up to 5x faster throughput than comparable open models. It was trained on roughly 26 trillion tokens before undergoing post-training with 3.8 million samples and reinforcement learning in NeMo Gym.

For local setups, Bolt 74B-A9B uses native 2-bit quantization to pack an agent-focused model into just 24GB of memory. It lets you offload multi-step workflows like web research, data extraction, and code debugging directly to your local machine. It currently holds the number one open-model spot on PinchBench and ranks second on BrowseComp for browsing agents.

Hardware specs and where to get it

Running Super VL locally requires serious gear. While active parameters are modest, the 120B footprint comes in at roughly 75GB on disk, demanding DGX systems, RTX PRO rigs, or enterprise GPU clusters. Third-party tests show it screaming along at over 600 tokens per second, making it built for high-concurrency pipelines.

Bolt 74B-A9B is far more accessible. It runs comfortably on standard 24GB GeForce RTX desktop GPUs and workstation setups, with out-of-the-box support across llama.cpp, Ollama, LM Studio, Unsloth, and vLLM. A smaller 4GB variant, Bolt E4B, is scheduled to drop on October 23.

The model weights, training recipes, and open datasets covering tool use and 40 vision-language benchmarks are available on Hugging Face and ModelScope, as well as production-ready NVIDIA NIM microservices.

Why it matters

If you build local workflows or operate privacy-sensitive internal tools, Bolt 74B-A9B is well worth testing today. Up until now, fitting capable agentic models into 24GB VRAM required heavy compromises; native 2-bit quantization makes local, untethered web and coding agents genuinely practical on desktop hardware.

Keep expectations calibrated on raw smarts, though. Independent testers note that Super VL lags behind models like Qwen 3.8-27B and MiniMax on pure benchmark reasoning. If your task requires solving convoluted math puzzles, look elsewhere. But if your team needs to ingest millions of PDFs, extract structured JSON, and trigger downstream tools at extreme speed and minimal cost, this architecture is a massive practical win.