Llama 4 · released 2025-04 · 402B parameters (17.2B active per token)
MoE on every other layer (24 MoE, 24 dense with 16384 FFN). Chunked local attention (8192) on 3 of 4 layers. Official FP8 checkpoint also published. params include ~0.87B vision encoder.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.75 GB | 0.38 GB |
| 8,192 tokens | 1.50 GB | 0.75 GB |
| 32,768 tokens | 2.63 GB | 1.31 GB |
| 131,072 tokens | 7.13 GB | 3.56 GB |
| 1,048,576 tokens | 49.1 GB | 24.6 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 748.0 GB |
| FP8 (W8A8) | 8 | 377.9 GB |
| INT8 (W8A8 / W8A16) | 8 | 377.9 GB |
| NVFP4 | 4.5 | 215.9 GB |
| INT4 AWQ / GPTQ | 4.25 | 204.4 GB |
| GGUF Q8_0 | 8.5 | 397.4 GB |
| GGUF Q6_K | 6.56 | 306.7 GB |
| GGUF Q5_K_M | 5.69 | 266.0 GB |
| GGUF Q4_K_M | 4.89 | 228.6 GB |
| GGUF Q3_K_M | 3.91 | 182.8 GB |
| GGUF Q2_K | 3.35 | 156.6 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) | 4× · 187 tok/s | 4× · 86.5 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 165 tok/s | 4× · 60.8 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 8× · 106 tok/s | 8× · 59.2 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 4× · 170 tok/s | 4× · 65.3 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 151 tok/s | 4× · 44.6 tok/s |
| L40S 48GBFP8 (W8A8) | 16× · 25.9 tok/s | 16× · 16.9 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 8× · 49.4 tok/s | 8× · 27.0 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 16× · 26.0 tok/s | 16× · 21.2 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 16× · 25.5 tok/s | 16× · 18.2 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 |