Nemotron · released 2025-04 · 253B parameters
NAS-pruned Llama-3.1-405B (DeciLM). 162 blocks: 64 have attention (8 KV heads; KV cache = 64 layers), 98 skip it; 81 blocks have an FFN (average width 53422; sum 4,327,168), 81 skip it. Counts from the safetensors index, verified against the exact parameter total.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 1.00 GB | 0.50 GB |
| 8,192 tokens | 2.00 GB | 1.00 GB |
| 32,768 tokens | 8.00 GB | 4.00 GB |
| 131,072 tokens | 32.0 GB | 16.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 472.0 GB |
| FP8 (W8A8) | 8 | 243.8 GB |
| INT8 (W8A8 / W8A16) | 8 | 243.8 GB |
| NVFP4 | 4.5 | 144.0 GB |
| INT4 AWQ / GPTQ | 4.25 | 136.9 GB |
| GGUF Q8_0 | 8.5 | 250.7 GB |
| GGUF Q6_K | 6.56 | 193.5 GB |
| GGUF Q5_K_M | 5.69 | 167.9 GB |
| GGUF Q4_K_M | 4.89 | 144.3 GB |
| GGUF Q3_K_M | 3.91 | 115.3 GB |
| GGUF Q2_K | 3.35 | 98.8 GB |
One server, 8K context, FP8 on data center GPUs and GGUF Q4_K_M on consumer and Apple hardware.
| GPU | 1 user | 32 users |
|---|---|---|
| B200 SXM (HGX B200)FP8 (W8A8) | 2× · 27.9 tok/s | 2× · 25.0 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 26.7 tok/s | 4× · 24.0 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 4× · 20.8 tok/s | 8× · 21.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 2× · 19.6 tok/s | 2× · 17.4 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 22.0 tok/s | 4× · 18.3 tok/s |
| L40S 48GBFP8 (W8A8) | 8× · 7.9 tok/s | 8× · 6.1 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 4× · 9.9 tok/s | 4× · 8.1 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 8× · 12.0 tok/s | 8× · 9.5 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 8× · 10.0 tok/s | 16× · 5.5 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |