Holotron-12B

<h1 style="display:flex; align-items:center; gap:20px; margin:0;"> <span>Holotron-12B &nbsp;&nbsp;&nbsp;&nbsp; - </span> <img src="https://cdn-uploads.huggingface.co/production/uploads/68c2d2a93ded6e08ec71ef8c/StqzozR1CHaw7ccLU5bfq.png" alt="Holotron logo" style="height: 50px; width: auto;" /> </h1> <!--- <h1 style="display:flex; align-items:center; gap:20px; margin:0;"> <span>Holotron-12B</span> <span>|</span> <img src="https://cdn-uploads.huggingface.co/production/uploads/68c2d2a93ded6e08ec71ef8c/StqzozR1CHaw7ccLU5bfq.png", alt="Holotron logo", height="1"/> </h1> --->

Model Overview

Holotron-12B is a high-throughput, multimodal Vision-Language Model (VLM) designed specifically as a policy model for computer-use agents.

Developed through a close collaboration between H Company and NVIDIA research labs, this model is post-trained from the open NVIDIA Nemotron Nano VL architecture on H Company’s proprietary data mixture. It is optimized for scale, production efficiency, and handling long contexts with multiple images in interactive environments.

• Developed by: H Company & NVIDIA
• Model Type: Multimodal Vision-Language Model (Hybrid SSM-Attention)
• Base Model: NVIDIA's Nemotron-Nano-12B-v2-VL-BF16
• Parameters: ~12B
• License: NVIDIA Open Model License
• Release Date: 16th March 2026
• Blog Post: https://hcompany.ai/holotron-12b

Intended use

Unlike general-purpose multimodal models optimized for static vision or simple instruction following, Holotron-12B is trained for multimodal agentic workloads.

It is designed to serve as the "brain" for agents that must:

• Perceive complex screens and UI elements.
• Decide on multi-step workflows.
• Act efficiently in interactive environments (Web, Desktop, Mobile).

Its high-throughput capabilities make it an ideal choice for data generation, annotation, and online reinforcement learning loops.

---

Training Strategy

Holotron-12B was trained in two stages. We started from Nemotron-Nano-12B-v2-VL-BF16, a multimodal base model published by NVIDIA. We then performed supervised fine-tuning on H Company’s proprietary localization and navigation data mixture, focusing on screen understanding, grounding, and UI-level interactions.

The final checkpoint was trained on approximately 14 billion tokens.

---

Architecture: Hybrid SSM

Holotron-12B utilizes a Hybrid State-Space Model (SSM) and attention mechanism. This architecture allows for superior scalability compared to pure transformer models.

Linear Complexity: Avoids the quadratic computation cost associated with full attention mechanisms.

Reduced Memory Footprint: SSMs act as a linear recurrent model, storing only a constant state per layer regardless of sequence length. This eliminates the massive KV Cache requirements of standard transformers for long sequences.

Throughput: Capable of maintaining high throughput even with long context histories and multiple high-resolution images.

---

Results

High throughput

On WebVoyager in a real-world multimodal agent setup (long context, multiple high-resolution images, concurrency up to 100):

• Single NVIDIA H100, vLLM with latest SSM optimizations (v0.14.1)
• Holotron-12B achieved >2x higher throughput vs Holo2-8B
• In a controlled setup, throughput scaled to ~8.9k tokens/s at concurrency = 100, while Holo2-8B plateaued around ~5.1k tokens/s

Navigation and Localization: Computer use

On computer-use and navigation benchmarks, Holotron-12B shows strong improvements over the Nemotron base model and strong performance with established agent models. WebVoyager performance increased from 35.1% to 80.5%, exceeding Holo2-8B’s performance on the benchmark and illustrating the model’s ability to perform in an agentic setting.

---

Get Started with the Model

Please refer to https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16 for getting started with this architecture.

Dependencies

pip install torch "transformers>4.53,<4.54" causal_conv1d timm "mamba-ssm==2.2.5" accelerate open_clip_torch numpy pillow

Conclusion

Holotron-12B demonstrates that the NVIDIA Nemotron VL model provides a strong foundation for real-world multimodal agents when paired with the right training setup and infrastructure work.

The model offers strong agent performance, significantly improved inference throughput, and a clear path for future improvements, particularly around higher-resolution vision training.

We look forward to seeing what others build with Holotron-12B.

Citation

@misc{hai2026holotron2,
      title={Holotron2}, 
      author={H Company},
      year={2026},
      url=https://huggingface.co/collections/Hcompany/Holotron-12B, 
}