---
title: "Holotron-12B - High Throughput Computer Use Agent"
description: "Holotron-12B is a multimodal computer-use model released by H Company and available on Hugging Face. Post-trained from NVIDIA’s Nemotron-Nano-12B-v2-VL-BF16 on proprietary data, it targets agents that must perceive, decide, and act in interactive environments while handling long contexts with multiple images. Its hybrid State-Space Model and attention architecture stores a constant state per layer per generated sequence instead of attention’s per-token key and value activations, reducing memory demands for long-context inference. In a WebVoyager evaluation on one H100 with vLLM, it reached 8.9k tokens/s at concurrency 100 versus 5.1k for Holo2-8B, and its benchmark score rose from 35.1% to 80.5% after training on localization and navigation data."
---

# Holotron-12B - High Throughput Computer Use Agent

[Hugging Face](https://yomu.fyi/company/hugging-face) · Pierre-Louis Cedoz, Hamza Benchekroun, Aurélien Lac, delfosse, Tony Wu, Antoine Bonnet, Kai Yuan, Aleix Cambray (H-AI), Alexandra, Axel Moyal · Mar 17, 2026

**Type:** Announcement

## Summary

Holotron-12B is a multimodal computer-use model released by H Company and available on Hugging Face. Post-trained from NVIDIA’s Nemotron-Nano-12B-v2-VL-BF16 on proprietary data, it targets agents that must perceive, decide, and act in interactive environments while handling long contexts with multiple images. Its hybrid State-Space Model and attention architecture stores a constant state per layer per generated sequence instead of attention’s per-token key and value activations, reducing memory demands for long-context inference. In a WebVoyager evaluation on one H100 with vLLM, it reached 8.9k tokens/s at concurrency 100 versus 5.1k for Holo2-8B, and its benchmark score rose from 35.1% to 80.5% after training on localization and navigation data.

## Context

H Company wanted a computer-use policy model that could scale efficiently in production, handle long contexts with multiple images, and perform well on agent benchmarks. The stated motivation includes throughput-bound workloads such as data generation, annotation, and online reinforcement learning.

## Approach / What changed

The model was post-trained in two stages from NVIDIA’s Nemotron-Nano-12B-v2-VL-BF16. Supervised fine-tuning used H Company’s proprietary localization and navigation data mixture, focusing on screen understanding, grounding, and UI-level interactions; the final checkpoint used approximately 14 billion tokens. Its hybrid SSM-attention architecture was served with vLLM and SSM optimizations.

## Takeaways

- The hybrid SSM design stores a constant state per layer and generated sequence, unlike attention’s per-token KV cache, reducing memory use as context length grows.
- At concurrency 100 on a single H100, Holotron-12B reached 8.9k total tokens per second, while Holo2-8B reached 5.1k tokens per second in the controlled comparison.
- WebVoyager performance improved from 35.1% for the Nemotron base model to 80.5% for Holotron-12B; the model also improved on OS-World-G, GroundUI, and WebClick.

**Tags:** [LLMs](https://yomu.fyi/topic/llm), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Performance](https://yomu.fyi/topic/performance)

- Source: [Hugging Face](https://huggingface.co/blog/Hcompany/holotron-12b)
- Source URL: https://huggingface.co/blog/Hcompany/holotron-12b
- Ingested by Yomu: 2026-08-29T01:46:34.205Z

[Read original post](https://huggingface.co/blog/Hcompany/holotron-12b)