← Back
Hugging Face
H Company releases Holotron-12B, 12B-parameter computer-use model with SSM architecture for efficient multimodal agent inference
· releasemodelfeatureintegrationperformance · huggingface.co ↗

Model Overview

H Company has released Holotron-12B, a multimodal computer-use agent model designed for efficient production deployment. The model is post-trained from NVIDIA's open Nemotron-Nano-2 VL foundation and is now available on Hugging Face as part of H Company's collaboration with the NVIDIA Inception Program.

Key Technical Innovations

Unlike most multimodal models that optimize for static vision or instruction-following, Holotron-12B is purpose-built as a policy model for agentic systems that must perceive, decide, and act in real-time interactive environments. The model's architecture combines a hybrid State-Space Model (SSM) with attention mechanisms—a departure from purely transformer-based designs.

Inference Efficiency Advantages

The SSM hybrid architecture delivers significant memory and throughput benefits:

  • Reduced KV Cache: Traditional transformers store key-value activations per token and layer, creating memory bottlenecks. SSMs use a linear recurrent design with only constant state per layer, independent of sequence length
  • Long-Context Support: The architecture efficiently handles multiple high-resolution images and extended interaction histories without quadratic computational overhead
  • Production Scalability: Optimized for high-throughput serving at inference time

Performance & Use Cases

The model excels on real-world agentic workloads including the WebVoyager Benchmark, which features long context windows, multiple images, and complex interaction sequences. This makes Holotron-12B particularly suitable for:

  • Browser automation and web navigation agents
  • Desktop/GUI interaction tasks
  • Multi-step computer use scenarios requiring image understanding

Getting Started

The model is available on Hugging Face and ready for production deployment. Developers can leverage Holotron-12B for building computer-use agents that require efficient inference at scale while maintaining strong performance on multimodal reasoning tasks.