Gpt Oss 20b
openai · released 2025-08-05 · apache-2.0 license
gpt-oss-20b is OpenAI's small open-weight reasoning MoE (21B total, 3.6B active, Aug 2025, Apache-2.0): 91.7% AIME 2025, 71.5% GPQA, 60.7% SWE-bench Verified (high reasoning, no tools). Artificial Analysis index 24.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 21.51B total · MoE, — active |
| Architecture | gpt_oss |
| Context window | 131K tokens |
| Knowledge cutoff | 2024-06-30 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 71.5% |
|---|---|
| SWE-bench Verified | 60.7% |
| AIME | 91.7% |
| MMLU-Pro | 85.3% |
| BFCL v3 (tool use) | 54.8% |
| Composite score | 6.24 |
| Community rating | 5.0★ (1 reviews, 0 net votes) |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 14 GB | 12.5 GB | — | 131K |
Strengths & weaknesses
Strengths: Strong math/reasoning for a small open model (21B total, 3.6B active); Apache-2.0, fast and cheap, 131K context; Leading intelligence among similar-size open models
Weaknesses: Knowledge-intensive tasks (GPQA, HLE) lag larger models; Text-only