OpenAI
GPT-OSS 20B
Frontier8.1MDownloads4.7KCurtidasAug 2025Publicado128K tokensContextoApache 2.0Licença90 ForteQualidade
GPT-OSS 20B (21B parameters) requires approximately 17.1 GB of VRAM with Q4_K_M quantization. As a Mixture of Experts model with 3.5999999046325684B active parameters, it uses less memory than its total parameter count suggests. For the best balance of quality and speed, we recommend hardware with at least 20 GB of VRAM.
Comece agora
— copie e cole para rodar localmenteCopy-paste commands to run GPT-OSS 20B on your machine.
Run
ollama run gpt-ossQuick specs
Parameters21B (3.6B active)
Architecturemoe (MoE)
Context128K tokens
Modalitytext
Min RAM8.2 GB
Rec. RAM12.8 GB (Q4_K_M)
LicenseApache 2.0
FamilyGPT-OSS
✓ Chat✓ Reasoning
About this model
- •OpenAI's first open-weight model under Apache 2.0 license
- •MoE architecture: 24 layers, 32 experts, top-4 routing per token
- •Configurable reasoning effort: low, medium, and high modes
- •Fits in 16GB VRAM with MXFP4 quantization
Modelos relacionados
Escolhas rápidas
Melhor hardware
Melhores opções para GPT-OSS 20B
Rodar este modelo
Opções de quantização
Estimativas de VRAM por nível de quantização
No hardware detected — fit column shows raw VRAM estimates
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.2 GB | Low | — |
Q3_K_S | 3 | 10.3 GB | Low | — |
NVFP4 | 4 | 11.8 GB | Medium | — |
Q4_K_M | 4 | 12.8 GB | Medium | — |
Q5_K_M | 5 | 15.1 GB | High | — |
Q6_K | 6 | 17.2 GB | High | — |
Q8_0 | 8 | 22.5 GB | Very High | — |
F16 | 16 | 43.1 GB | Maximum | — |
Quality benchmarks
GPT-OSS 20B benchmark scores
Coding
SWE-bench Verified60.7%
HumanEval+—
Aider Polyglot—
LiveCodeBench74.6%
Reasoning
MMLU-Pro—
GPQA Diamond71.5%
MATH-500—
ARC Challenge—
Source: official · 2025-08-15
Compatibilidade de hardware
Estimativas de compatibilidade para todo o hardware
Computing compatibility...
Detalhamento de memória
Reference: RTX 2060 6GB
Weights12.8 GB
KV Cache2.4 GB
Runtime1.2 GB
Headroom0.6 GB
Perguntas frequentes
FAQ — GPT-OSS 20B
Veja também