gpt-oss · released 2025-08 · 20.9B parameters (3.61B active per token)
Params per OpenAI model card (20.91B total / 3.61B active; HF metadata shows 21.51B because packed MXFP4 scales are counted). 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 | 0.10 GB | 0.05 GB |
| 8,192 tokens | 0.19 GB | 0.10 GB |
| 32,768 tokens | 0.75 GB | 0.38 GB |
| 131,072 tokens | 3.00 GB | 1.50 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 38.9 GB |
| FP8 (W8A8) | 8 | 21.6 GB |
| INT8 (W8A8 / W8A16) | 8 | 21.6 GB |
| NVFP4 | 4.5 | 14.1 GB |
| INT4 AWQ / GPTQ | 4.25 | 13.5 GB |
| GGUF Q8_0 | 8.5 | 20.7 GB |
| GGUF Q6_K | 6.56 | 16.0 GB |
| GGUF Q5_K_M | 5.69 | 13.9 GB |
| GGUF Q4_K_M | 4.89 | 11.9 GB |
| GGUF Q3_K_M | 3.91 | 9.52 GB |
| GGUF Q2_K | 3.35 | 8.15 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× · 461 tok/s | 1× · 142 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 346 tok/s | 1× · 90.8 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 273 tok/s | 1× · 65.3 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 368 tok/s | 1× · 99.2 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 265 tok/s | 1× · 59.6 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 90.1 tok/s | 1× · 17.8 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 154 tok/s | 1× · 32.3 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 192 tok/s | 1× · 49.5 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 133 tok/s | 1× · 29.3 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 96.5 tok/s | 1× · 19.5 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 30.5 tok/s | 1× · 5.7 tok/s |