← All models

Nemotron-H 8B Reasoning 128K

NVIDIA · 2025-06-06 · 8.1B parameters

Nemotron-H 8B Reasoning 128K is NVIDIA's open-weight 8B hybrid Mamba-Transformer reasoning model post-trained from Nemotron-H 8B Base and released on June 6, 2025; the BF16 checkpoint contains exactly 8,100,852,736 parameters. It was trained with staged supervised fine-tuning on reasoning and non-reasoning data, long-context extension to a 131,072-token context, GRPO, and checkpoint interpolation, and reasoning can be forced on or off with a system-prompt tag. The weights are released under the NVIDIA Internal Scientific Research and Development Model License.

Benchmark scores