Holotron-12B: Fast Multimodal Agent for Production Computer Use

H Company has unveiled Holotron-12B, a multimodal model purpose-built for computer-use agents that need to perceive, decide, and act in interactive environments at scale. The 12-billion parameter model, available on Hugging Face, is post-trained from NVIDIA's open Nemotron-Nano-2 VL base model using proprietary data.

Architecture & Inference Speed#

The breakthrough lies in Holotron-12B's hybrid State-Space Model (SSM) and attention architecture, which trades the quadratic memory cost of pure transformers for linear-recurrent inference. Instead of storing K and V cache activations per token and layer—a bottleneck for long-context work—SSMs maintain only a constant state per layer, regardless of sequence length.

On real-world multimodal workloads, the efficiency gains are substantial. Tested on the WebVoyager Benchmark with 100 concurrent workers, multiple high-resolution images, and long interaction histories, Holotron-12B achieved over 2x higher throughput than Holo2-8B on a single H100 GPU running vLLM. The model sustained 8.9k tokens/second at peak concurrency, while Holo2-8B plateaued at 5.1k tokens/second—a gap the team attributes to the Nemotron architecture's superior VRAM utilization.

Agentic Performance#

Holotron-12B was trained in two stages: supervised fine-tuning on H Company's localization and navigation data (∼14 billion tokens total), focusing on screen understanding and UI-level interactions. The results are strong across benchmarks:

• WebVoyager: 80.5% (up from 35.1% on the base Nemotron model, and exceeding Holo2-8B)
• Localization benchmarks: OS-World-G, GroundUI, and WebClick all show substantial improvements over the base model

The architecture makes Holotron-12B particularly suited for data generation, annotation, and online reinforcement learning—workloads where throughput and batch efficiency are critical.

Looking Ahead#

H Company is preparing post-trained variants of NVIDIA's newly announced Nemotron 3 Omni to push reasoning and multimodal precision further, targeting enterprise-scale autonomous computer-use deployments.

The model is licensed under the NVIDIA Open Model License.