OLMo 2 · released 2024-11 · 13.7B parameters
Memory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 3.13 GB | 1.56 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 25.5 GB |
| FP8 (W8A8) | 8 | 14.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 14.7 GB |
| NVFP4 | 4.5 | 9.94 GB |
| INT4 AWQ / GPTQ | 4.25 | 9.60 GB |
| GGUF Q8_0 | 8.5 | 13.6 GB |
| GGUF Q6_K | 6.56 | 10.5 GB |
| GGUF Q5_K_M | 5.69 | 9.09 GB |
| GGUF Q4_K_M | 4.89 | 7.81 GB |
| GGUF Q3_K_M | 3.91 | 6.24 GB |
| GGUF Q2_K | 3.35 | 5.35 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× · 178 tok/s | 1× · 34.2 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 117 tok/s | 1× · 20.9 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 85.8 tok/s | 2× · 28.1 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 128 tok/s | 1× · 23.0 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 74.3 tok/s | 2× · 18.3 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 24.1 tok/s | 4× · 13.6 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 43.4 tok/s | 2× · 13.5 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 62.4 tok/s | 8× · 29.1 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 37.9 tok/s | 8× · 20.5 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 25.5 tok/s | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 7.8 tok/s | > 16 |