From the source
H Company published a research note detailing the post-training of
NVIDIA's Nemotron 3.5 Super VL 120B-A12B vision model for computer use.
Their resulting checkpoint, Holotron4-Super-Preview, improved on the base model across benchmarks including OSWorld (+25.1), OSWorld 2.1 (+12.0), and AutomationBench (+17.9).
The recipe involved supervised fine-tuning, weight merging with the base model, and two reinforcement-learning experts merged back in.





