Nemotron · released 2025-07 · 49.9B parameters
NAS-pruned Llama-3.3-70B (DeciLM). Irregular blocks: only 49 of 80 layers have attention (8 KV heads, so KV cache = 49 layers); the other 31 skip attention. All 80 layers keep an FFN of varying width; ffn = average 20531 (sum 1,642,496, i.e. mults 0.5-5.25 x 8192). Counts from model.safetensors.index.json, verified against the exact parameter total. v1 (Mar 2025) has the same architecture.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.77 GB | 0.38 GB |
| 8,192 tokens | 1.53 GB | 0.77 GB |
| 32,768 tokens | 6.13 GB | 3.06 GB |
| 131,072 tokens | 24.5 GB | 12.3 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 92.9 GB |
| FP8 (W8A8) | 8 | 50.4 GB |
| INT8 (W8A8 / W8A16) | 8 | 50.4 GB |
| NVFP4 | 4.5 | 31.7 GB |
| INT4 AWQ / GPTQ | 4.25 | 30.4 GB |
| GGUF Q8_0 | 8.5 | 49.3 GB |
| GGUF Q6_K | 6.56 | 38.1 GB |
| GGUF Q5_K_M | 5.69 | 33.0 GB |
| GGUF Q4_K_M | 4.89 | 28.4 GB |
| GGUF Q3_K_M | 3.91 | 22.7 GB |
| GGUF Q2_K | 3.35 | 19.4 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) | 1× · 70.2 tok/s | 1× · 49.9 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 44.3 tok/s | 1× · 31.1 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 31.7 tok/s | 2× · 39.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 48.6 tok/s | 1× · 34.1 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 31.6 tok/s | 1× · 18.4 tok/s |
| L40S 48GBFP8 (W8A8) | 2× · 15.8 tok/s | 2× · 10.6 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 15.6 tok/s | 1× · 10.7 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 2× · 39.5 tok/s | 4× · 28.7 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 2× · 25.7 tok/s | 4× · 19.7 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 10.0 tok/s | 1× · 2.0 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 2.7 tok/s | 1× · 1.9 tok/s |