• Sources: primary
  • Summary: The report describes distillation, long-context training, native speculative decoding, agentic software-engineering mid-training and multi-teacher on-policy distillation for Super. NVIDIA reports 51.78% on SWE-bench Pro for Super, NVFP4 quantization recovering 99.62% median accuracy versus BF16, and up to 4,856 output tokens per second per GPU on GB200 at 50 or more tokens per second per user. Benchmark results are NVIDIA's own.
  • Why it matters: It documents a full open-model recipe, including NVFP4 deployment and MTP speculative decoding gains, that teams serving MoE models on Blackwell can compare against.