Holotron-12B: A Production-Ready Computer-Use Model

H Company has released Holotron-12B, a multimodal model specifically designed for computer-use agents that need to perceive, decide, and act efficiently in interactive environments. The model is now available on Hugging Face and represents a collaboration between H Company and NVIDIA, with H Company as part of the NVIDIA Inception Program.

Unlike most multimodal models that optimize for static vision tasks or instruction-following, Holotron-12B is purpose-built as a policy model for agentic workloads. The model was post-trained from NVIDIA's open Nemotron-Nano-2 VL foundation using H Company's proprietary data mixture, demonstrating significant performance improvements through additional training on the base architecture.

Hybrid SSM Architecture Enables Efficient Inference

The key innovation in Holotron-12B is its hybrid State-Space Model (SSM) and attention mechanism, which dramatically improves inference efficiency compared to purely transformer-based models. State-space models avoid the quadratic computation costs of full attention mechanisms, offering superior scalability for long-context inference—a critical requirement for agentic tasks involving multiple high-resolution images and extended interaction histories.

The architectural advantages are particularly pronounced in memory efficiency: while standard transformer attention must maintain K and V cache activations per token and layer, SSMs use only a constant state per layer regardless of sequence length. This linear memory footprint makes Holotron-12B significantly more suitable for production deployments handling long contexts.

Performance and Availability

On the WebVoyager Benchmark—a real-world multimodal agentic workload featuring long context, multiple high-resolution images, and complex interactions—Holotron-12B demonstrates strong performance. The model is optimized for both inference speed and quality, addressing a key gap in the market for production-grade computer-use agents.

Developers can access Holotron-12B on Hugging Face immediately and integrate it into agentic applications requiring efficient perception, decision-making, and action capabilities.