vLLM and Red Hat AI release FP8 and DSpark checkpoints for Nemotron 3.5 Super VL

vLLM and Red Hat AI release FP8 and DSpark optimized checkpoints for NVIDIA Nemotron 3.5 Super VL, a 120B multimodal mixture-of-experts model, on Hugging Face.

Red Hat AI says the two optimizations, run on 4× H100 GPUs with vLLM, together outperform either one alone on HumanEval, math reasoning, and RAG workloads. FP8 quantization improves throughput under load, and DSpark speculative decoding raises per-user streaming speed. vLLM says both checkpoints are open and ready to use on Hopper GPUs. The work used LLM Compressor, Speculators, and vLLM.