Nemotron 3 Super splits vLLM, llama.cpp

This is probably a backend-fidelity problem, not a temperature problem. NVIDIA's reference setup is much more specific than a bare server launch, so a community GGUF path can drift enough to move scores.

• –NVIDIA's quick start already standardizes `temperature=1.0` and `top_p=0.95`, and exposes `enable_thinking` through the chat template, so the knobs the poster tried are unlikely to explain a 15-point swing.
• –The official vLLM recipe adds a custom `super_v3` reasoning parser plus `--enable-expert-parallel`, `--enable-chunked-prefill`, and Mamba cache tuning, which suggests model-specific serving logic beyond generic inference defaults.
• –NVIDIA publishes BF16, FP8, and NVFP4 checkpoints with official vLLM, SGLang, TRT-LLM, and NIM guidance, while the llama.cpp run here depends on a community GGUF conversion, so this is not an apples-to-apples comparison.
• –The Gemma 3 sanity check, plus Reddit speculation about tokenizer/template mismatch, points more toward prompt serialization or reasoning-token handling than raw quantization quality.