Runs on ≤16 GB · Hugging Face ↗
Not independently scored by Artificial Analysis — its base model nvidia/NVIDIA-Nemotron-Nano-12B-v2-Base is the closest proxy
NVIDIA's open 9B on the Nemotron-H hybrid architecture, with a reasoning mode. Free on OpenRouter's :free routes, and small enough for any card.
BigCodeBench-Hard pass@1 22% (27/121), via OpenRouter :free route (hosted) — comparable full-148. The brutal counterpart to HumanEval — where HumanEval saturates near the top, BCB-Hard spreads the field, so this is the number that actually separates coding ability. A modest 9B on hard problems — fine as a fast free draft, not a top coder.
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈8 GB at 32K, fp16).
| Quantization | Size |
|---|---|
| Q2_K | 5.0 GB |
| Q4_0 | 5.3 GB |
| Q4_1 | 5.8 GB |
| Q4_K_S | 6.2 GB |
| Q4_K_M | 6.5 GB |
| Q5_K_M | 7.1 GB |
| Q6_K | 9.1 GB |
| Q8_0 | 9.5 GB |
| F16 | 17.8 GB |
~6 GB at Q4 — fits an 8 GB card. Free on OpenRouter :free, which is how we measured it.