H Company just dropped Holotron-12B, and the numbers demand attention. Post-trained from NVIDIA’s Nemotron-Nano-2 VL model, this 12-billion-parameter multimodal agent scored 80.5% on the WebVoyager benchmark—more than doubling its base model’s 35.1%. It even edged out H Company’s own Holo2-8B.
The real story isn’t just the score. It’s the engine underneath. Holotron-12B runs on a hybrid State-Space Model (SSM) and attention architecture inherited from Nemotron. That means it sidesteps the quadratic compute costs and memory bloat of a standard transformer’s KV cache. SSMs store only a constant state per layer, so memory footprint stays flat regardless of how many images or interaction steps pile up.
How flat? In a controlled test with 100 concurrent workers on a single H100 GPU, Holotron-12B hit 8.9k tokens per second. Holo2-8B plateaued at 5.1k. That’s a 2x throughput advantage under load. For teams running data generation, annotation, or online RL pipelines, that’s not a nice-to-have. It’s the difference between a batch job finishing overnight or bleeding into the weekend.
H Company trained the model in two stages: starting from Nemotron-Nano-12B-v2-VL-BF16, then fine-tuning on roughly 14 billion tokens of proprietary localization and navigation data. The focus was screen understanding, grounding, and UI-level interactions. The results show up on localization benchmarks too—OS-World-G, GroundUI, and WebClick all got substantial lifts. The model is available now on Hugging Face under NVIDIA’s open license, and with Nemotron 3 Omni already announced, H Company is betting this architecture has legs.
