Nemotron-3.5-Super-Puzzletron-VL-123B-A14B-NVFP4
Super-Puzzletron is built on NVIDIA Nemotron 3.5 Super and post-trained with 8 steps of GRPO on an exact-answer word-puzzle task. It doubles the base model's accuracy on its training task and improves on unseen puzzles. It is a standard NVFP4 checkpoint: stock vLLM serves it with no plugins or custom code.
Results
Procedurally generated word puzzles in a NeMo Gym environment, exact answers only. The base model and Super-Puzzletron were served by the same stock vLLM build in NVFP4.
| Base | Super-Puzzletron | Change | |
|---|---|---|---|
| Training task | 43.8% | 87.5% | 2.0× |
| Unseen puzzles | 58.3% | 64.6% | +11% relative |
| Unseen, level 1 | 65.6% | 78.1% | +12.5 points |
| Unseen, level 2 | 57.8% | 60.9% | +3.1 points |
| Unseen, level 3 | 51.6% | 54.7% | +3.1 points |
Usage
vllm serve LLMWildling/Nemotron-3.5-Super-Puzzletron-VL-123B-A14B-NVFP4 \
--tensor-parallel-size 2 \
--trust-remote-code \
--max-model-len 32768
Tested with vLLM 0.32, no plugins. For reasoning and tool-call parsing, use the options from the base model card.
License
OpenMDW-1.1, inherited from NVIDIA Nemotron 3.5 Super VL.