Holotron-12B High Throughput Computer Use Agent

H Company has released Holotron-12B, a multimodal model designed specifically as a policy model for computer-use agents. By post-training the NVIDIA Nemotron-Nano-2 VL model on a proprietary data mixture, H Company has developed an agent that can perceive, decide, and act efficiently within interactive environments, optimizing for production scale and performance.

High Throughput via Hybrid SSM Architecture

Holotron-12B achieves high inference efficiency by utilizing a hybrid State-Space Model (SSM) and attention mechanism. This architecture is specifically optimized for high-throughput serving, providing several advantages over purely transformer-based models:

  • Reduced Memory Footprint: Unlike vanilla attention, which requires storing KV activations per token and layer (the KV Cache), SSMs are linear recurrent models that store only a constant state per layer per generated sequence, regardless of sequence length.
  • Linear Scalability: The design avoids the quadratic computation costs associated with full attention, making it more efficient for long-context inference involving multiple high-resolution images and extensive interaction histories.
  • Increased Throughput: In tests using the WebVoyager Benchmark on a single H100 GPU with vLLM (v0.14.1), Holotron-12B achieved over 2x higher throughput than Holo2-8B.

In controlled experiments, Holotron-12B's total token throughput rose steadily to 8.9k tokens/s at a maximum concurrency of 100, whereas Holo2-8B plateaued at 5.1k tokens/s. This demonstrates more effective VRAM utilization and a smaller memory footprint, allowing for larger effective batch sizes on the same hardware.

Training Methodology and Data

Holotron-12B was developed in two stages, starting from the NVIDIA Nemotron-Nano-12B-v2-VL-BF16 multimodal base model. The model underwent supervised fine-tuning on H Company’s proprietary localization and navigation data mixture, which focused on three primary areas:

  1. Screen understanding
  2. Grounding
  3. UI-level interactions

The final checkpoint was trained on approximately 14 billion tokens.

Performance Benchmarks

Holotron-12B shows significant improvements over the base Nemotron model and competitive performance against established agent models across several key benchmarks:

Agentic Performance

On the WebVoyager benchmark, Holotron-12B’s performance increased from 35.1% (base Nemotron) to 80.5%, exceeding the performance of Holo2-8B.

Localization and Grounding

The model also demonstrated substantial improvements over the base Nemotron model on localization and grounding benchmarks, including:

  • OS-World-G
  • GroundUI
  • WebClick

Future Outlook: Nemotron 3 Omni

Building on the results of Holotron-12B, H Company is preparing to post-train the next generation of multimodal models using the Nemotron 3 Omni architecture. This next iteration will leverage enhanced hybrid SSM-Attention and Mixture-of-Experts (MoE) foundations to improve reasoning capabilities, multimodal precision, and provide the low-latency performance required for massive-scale autonomous computer-use deployments.

Sources