Holotron-12B speeds up AI agents on the web
H Company has released Holotron-12B, a multimodal model designed to help AI agents use computers faster. In tests with one H100 GPU, it reached 8,900 tokens per second and raised its WebVoyager score from 35.1% to 80.5%.
Holotron-12B is a new AI model that can interpret a screen, decide what to do and operate digital interfaces. H Company has released it on Hugging Face to help computer-using agents run faster while consuming less memory.
The model is based on Nemotron-Nano-12B-v2-VL-BF16, NVIDIA's multimodal model for processing text and images. H Company then trained it on around 14 billion tokens and proprietary data covering localization, navigation, screen understanding, and interaction with buttons and other interface elements.
More speed when many agents are running
Holotron-12B's main difference is not only what it understands, but how it processes requests. Its architecture combines attention with a state space model, known as an SSM, a system that maintains a summarized state of the conversation instead of storing every detail from each generated token.
That reduces memory use when the agent works with long histories, many screenshots and several tasks at once. In practice, it allows more agents to run on the same hardware.
In a test using the WebVoyager benchmark, which measures the ability to navigate websites, H Company ran Holotron-12B on a single NVIDIA H100 GPU, with vLLM 0.14.1 and up to 100 simultaneous workers. The model reached 8,900 tokens per second, compared with 5,100 for Holo2-8B in the same experiment.
That represents more than twice the performance of Holo2-8B, according to the company. The advantage is especially relevant for tasks that require generating many examples, labeling data or training agents through continuous interaction with digital environments.
It also improves accuracy
Speed alone is not useful if the agent clicks in the wrong place. In WebVoyager, Holotron-12B raised the original Nemotron model's score from 35.1% to 80.5%, while also outperforming Holo2-8B in the test.
It also recorded improvements in localization and interface understanding evaluations such as OS-World-G, GroundUI and WebClick. These tests measure whether the model correctly identifies an element on screen and performs the appropriate action, such as selecting a menu, filling in a field or clicking a button.
Holotron-12B is optimized to perceive, decide and act in interactive environments.
What this means for you
The model is not a general-purpose assistant designed mainly for conversation. It is built as the component that controls an agent capable of using a computer, which can be applied to:
- Automating repetitive tasks on websites.
- Testing applications and detecting interface bugs.
- Generating data to train other agents.
- Processing large volumes of navigation or annotation.
Because it is available on Hugging Face under an open NVIDIA license, companies and researchers can test and adapt it, although they will need powerful infrastructure to run it effectively. The published results also come from specific conditions: one H100 GPU, vLLM and concurrency of up to 100 requests.
H Company is already preparing a version based on Nemotron 3 Omni, NVIDIA's newly announced multimodal family. The next step will be to see whether this combination of lower memory use and better navigation can hold up in real commercial applications, where the speed, accuracy and safety of every action all matter.