Model notes
OpenReasoning Nemotron 1.5B
Small dense Nemotron reasoning model built on the Qwen2.5 1.5B geometry, aimed at strong math and code behavior on modest hardware.
1.5B dense • 32,768 context • 2 KV heads
Architecture
Model spec
Architecture
Total params
Active params
Layers
Hidden size
Attention heads
KV heads
KV-bearing layers
Context length
Modality
License
Why it matters
Why memory behaves this way
Research highlight
NVIDIA post-trains the Qwen2.5 1.5B base for reasoning while keeping the dense grouped-query architecture intact, so the memory geometry stays predictable.
Memory note
This behaves like a classic dense Qwen2.5-style checkpoint where resident weights dominate and KV cache follows the standard grouped-attention path.
Checkpoints
Official profiles
Official BF16 checkpoint
BF16 checkpoint
NVIDIA ships OpenReasoning-Nemotron-1.5B in Hugging Face Transformers format, and v1 models it as a standard dense Qwen2.5-derived checkpoint across the supported runtimes.
Sources