← All models

Nemotron Cascade 8B

NVIDIA · 2025-12-08 · 8.2B parameters

Nemotron Cascade 8B is NVIDIA's open-weight general-purpose model post-trained from Qwen3-8B-Base with sequential, domain-wise reinforcement learning (Cascade RL) after SFT; a single checkpoint operates in both thinking and instruct (non-reasoning) modes. The BF16 checkpoint contains 8,190,735,360 parameters in the Qwen3 dense architecture (36 layers, GQA) with a 32,768-token native context that NVIDIA extends to 64K with YaRN for evaluation. The card documents a December 8, 2025 release, although the Hugging Face repository was created on December 16, 2025.

Benchmark scores