Olmo 3 · released 2025-12 · 32.2B parameters
3 sliding-window (4096) : 1 full attention. Olmo 3 32B Think (Nov 2025) has the same architecture.
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 | 1.25 GB | 0.63 GB |
| 32,768 tokens | 2.75 GB | 1.38 GB |
| 65,536 tokens | 4.75 GB | 2.38 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 60.0 GB |
| FP8 (W8A8) | 8 | 31.9 GB |
| INT8 (W8A8 / W8A16) | 8 | 31.9 GB |
| NVFP4 | 4.5 | 19.6 GB |
| INT4 AWQ / GPTQ | 4.25 | 18.8 GB |
| GGUF Q8_0 | 8.5 | 31.9 GB |
| GGUF Q6_K | 6.56 | 24.6 GB |
| GGUF Q5_K_M | 5.69 | 21.4 GB |
| GGUF Q4_K_M | 4.89 | 18.3 GB |
| GGUF Q3_K_M | 3.91 | 14.7 GB |
| GGUF Q2_K | 3.35 | 12.6 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) | 1× · 105 tok/s | 1× · 70.2 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 67.5 tok/s | 1× · 44.1 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 48.7 tok/s | 1× · 31.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 73.7 tok/s | 1× · 48.3 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 49.7 tok/s | 1× · 25.9 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 13.3 tok/s | 2× · 15.3 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 24.2 tok/s | 1× · 15.4 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 38.2 tok/s | 2× · 32.0 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 22.8 tok/s | 4× · 27.1 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 15.2 tok/s | 1× · 3.2 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 4.3 tok/s | 1× · 2.7 tok/s |