Runs on ≤32 GB · Hugging Face ↗
Coding index 25.3 (Artificial Analysis)
A newer NVIDIA 30B-total MoE (~3B active) using cascade routing — fast decoding for its capability, tuned for the NVIDIA stack.
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈2 GB at 32K, fp16).
| Quantization | Size |
|---|---|
| IQ4_XS | 18.2 GB |
| Q4_0 | 18.3 GB |
| Q3_K_M | 19.1 GB |
| Q4_1 | 20.1 GB |
| Q4_K_S | 22.4 GB |
| Q4_K_M | 24.7 GB |
| Q4 | 24.9 GB |
| Q5_K_M | 26.2 GB |
| Q8_0 | 33.6 GB |
| Q2_K | 36.5 GB |
| BF16 | 63.2 GB |
| Q6_K | 67.1 GB |
~25 GB at Q4 — a 32 GB card, or a 24 GB one with CPU offload of idle experts. vLLM / TensorRT-LLM give the best throughput on NVIDIA hardware.