H Company releases Holotron-12B — 2x throughput on H100 GPU

AI Impact Summary

H Company released Holotron-12B, a new multimodal model built on the NVIDIA Nemotron-Nano-2 VL architecture, optimized for high-throughput inference in production environments. The model’s hybrid SSM architecture and efficient VRAM utilization, demonstrated through benchmarks on the WebVoyager Benchmark, achieve over 2x throughput compared to Holo2-8B, making it suitable for data generation and online reinforcement learning workloads. This release represents a significant step towards scalable agentic intelligence.