Hermes Agent Adds NVIDIA Nemotron 3.5 Support
Nous Research integrated Hermes Agent with NVIDIA Nemotron 3.5 Super VL, reaching 63.44 on Hermes Bench and up to 5x higher throughput.
The multimodal Nemotron 3.5 Super VL scored 63.44 on Hermes Bench, delivering up to five times faster throughput than comparable local setups.
Nous Research has integrated Hermes Agent with NVIDIA’s newly launched Nemotron 3.5 open model family. The initial rollout focuses on Nemotron 3.5 Super VL, a multimodal model that posted a 63.44 score on the Hermes Bench evaluation. According to Nous Research, the architecture delivers token throughput up to five times faster than competing local options during autonomous agent workflows.
One Model for Vision, Tools, and Reasoning
Traditional visual agent architectures rely on a fractured pipeline: one model processes the screenshot or photo, turns that visual input into text, and feeds it to a separate language model to decide what to do next. The Nemotron 3.5 Super VL skips that split entirely. NVIDIA unified computer vision, logical reasoning, and tool calling inside a single set of weights.
Under the hood, Super VL uses a Mixture-of-Experts (MoE) design. While the model contains 120 billion total parameters, it routes tokens through only 12 billion active parameters at inference time. That sparse activation keeps compute manageable without sacrificing deep reasoning. NVIDIA also gave it a 1-million-token context window, tailored for long-running agents that accumulate massive session histories of tool outputs and terminal states.
In the NVIDIA NIM catalog, the quantized NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 build adds speculative decoding via Multi-Token Prediction (MTP), generating several tokens simultaneously to cut agent latency.
Bolt Models for Local Hardware
Running 120 billion parameters is out of reach for a standard developer rig. To cover local workstations and edge machines, NVIDIA released the Nemotron 3.5 Bolt family alongside the flagship.
The first available variant is Bolt 74B-A9B. Using native 2-bit quantization, it operates within 24 GB of VRAM, making it viable on high-memory GeForce RTX desktop cards and DGX Spark units.
For smaller setups, NVIDIA scheduled the Bolt E4B, a 4 GB footprint model, for release on October 23, 2026. The Hermes Agent team plans to package this lightweight variant directly into their desktop installer, though that depends on NVIDIA releasing the weights on schedule.
Setting Up Hermes Agent with NVIDIA NIM
Official Hermes Agent documentation already outlines how to hook into these models. Developers can direct calls to hosted NVIDIA NIM endpoints at build.nvidia.com or connect to on-premises enterprise setups such as DGX Station hardware.
There are two practical catches to keep in mind. First, unlike earlier Nous and NVIDIA partnerships, there is no free hosted tier on the Nous Portal this time around. You will need your own NVIDIA NIM API key or you will have to pull the open weights from Hugging Face onto your own iron.
Second, neither Nous Research nor NVIDIA released head-to-head Hermes Bench comparisons against closed frontier models like GPT-4o or Claude 3.5 Sonnet for this specific benchmark run. The throughput improvements over prior open local weights are clear, but how Super VL stacks up against leading commercial APIs in agentic reasoning remains an open question.