gpt-oss · released 2025-08 · 117B parameters (5.13B active per token)
Params per OpenAI model card (116.83B total / 5.13B active). Alternating sliding-window (128) / full attention with learned attention sinks; MoE experts stored in MXFP4 (attention, router, embeddings in BF16). YaRN x32 from 4096.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.15 GB | 0.07 GB |
| 8,192 tokens | 0.29 GB | 0.14 GB |
| 32,768 tokens | 1.13 GB | 0.56 GB |
| 131,072 tokens | 4.50 GB | 2.25 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 217.6 GB |
| FP8 (W8A8) | 8 | 111.0 GB |
| INT8 (W8A8 / W8A16) | 8 | 111.0 GB |
| NVFP4 | 4.5 | 64.3 GB |
| INT4 AWQ / GPTQ | 4.25 | 61.0 GB |
| GGUF Q8_0 | 8.5 | 115.6 GB |
| GGUF Q6_K | 6.56 | 89.2 GB |
| GGUF Q5_K_M | 5.69 | 77.4 GB |
| GGUF Q4_K_M | 4.89 | 66.5 GB |
| GGUF Q3_K_M | 3.91 | 53.2 GB |
| GGUF Q2_K | 3.35 | 45.6 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× · 385 tok/s | 1× · 50.5 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 285 tok/s | 1× · 31.1 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 2× · 279 tok/s | 2× · 41.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 303 tok/s | 1× · 34.2 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 236 tok/s | 1× · 23.1 tok/s |
| L40S 48GBFP8 (W8A8) | 4× · 104 tok/s | 4× · 19.8 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 2× · 156 tok/s | 2× · 20.1 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 4× · 114 tok/s | 4× · 44.3 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 4× · 104 tok/s | 4× · 29.3 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 72.4 tok/s | 1× · 6.3 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |