NVIDIA Nemotron Nano 9B V2
nvidia · released 2025-09-05 · other license
NVIDIA-Nemotron-Nano-9B-v2 (Aug 2025) is a 9B hybrid Mamba-Transformer reasoning model with strong math/coding (MATH-500 97.8, LiveCodeBench 71.1, AIME25 72.1) and toggleable reasoning. Sept-2024 cutoff.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 8.89B total |
| Architecture | nemotron_h |
| Context window | 131K tokens |
| Knowledge cutoff | 2025-03-31 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 64% |
|---|---|
| SWE-bench Verified | — |
| AIME | 72.1% |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | 5.43 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 6.7 GB | 5.2 GB | — | 131K |
Strengths & weaknesses
Strengths: Excellent math/coding for 9B (MATH-500 97.8, LiveCodeBench 71.1); Hybrid Mamba-2/Transformer, toggleable reasoning + thinking budget; Strong IFEval (90.3) and long context (RULER 78.9)
Weaknesses: Very low on hardest frontier QA (HLE 6.5); Restrictive NVIDIA Open Model License