@dair_ai
Great report from NVIDIA. They report that Nemotron 3.5 Super reaches 51.78% on SWE-bench Pro. That beats Nemotron 3 Ultra with about 4.6x fewer active parameters. Super has 120B parameters with 12B active, and it handles text, images, and video. Lightning is a 30B text model with 3B active. The technical report covers every stage of training. Super gets mid-training on agentic software engineering, then supervised fine-tuning, RL, and multi-teacher on-policy distillation. Serving numbers are also included. The stack reaches up to 4,856 output tokens/s/GPU on GB200 while holding at least 50 tokens/s/user. NVFP4 quantization recovers 99.62% median accuracy relative to BF16. https://t.co/Tt5etoqlpG Chat with Paper: https://t.co/Wjb3TfcsCn