← All models

Nemotron Cascade 14B Thinking

NVIDIA · 2025-12-08 · 14.8B parameters

Nemotron Cascade 14B Thinking is NVIDIA's open-weight reasoning-only model post-trained from Qwen3-14B-Base with SFT and sequential, domain-wise reinforcement learning (Cascade RL), released on December 8, 2025. The BF16 checkpoint contains 14,768,307,200 parameters in the Qwen3 dense architecture (40 layers, GQA) with a 32,768-token native context that NVIDIA extends with YaRN to 64K (90K for SWE-bench Verified) for evaluation.

Benchmark scores