Model

gpt-oss-20b (Medium)

OPEN

OpenAI · gpt-oss family · released Aug 5, 2025

21B MoE (3.6B active) open-weight sibling that runs on consumer hardware; Apache 2.0.

ReasoningCodingVisionFunction callingTool useAgentic

1367.6

Elo · rank #143

Parameters

21B

Active params

3.6B (MoE)

Context

128K tokens

Architecture

Mixture-of-experts transformer, 32 experts/layer, top-4 active, 24 layers

License

Apache 2.0

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

Aider PolyglotCoding26.6%#34

best: Claude Opus 4.5 · 89.4%

AIMEMath80.0%#78

best: GPT-5.2 · 100.0%

CodeforcesCoding1998#20

best: DeepSeek-V4-Pro (Think Max) · 3206

GPQA DiamondReasoning66.0%#139

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning7.0%#103

best: Claude Sonnet 5 · 57.4%

MMLUKnowledge84.0%#67

best: OpenAI o3 · 92.9%

MMMLUKnowledge73.5%#39

best: Gemini 3.1 Pro · 92.6%

SWE-bench VerifiedCoding53.2%#87

best: Claude Fable 5 · 95.0%

tau-benchAgents47.3%#23

best: Claude Opus 4.1 · 82.4%

Run it locally

VRAM @ Q4

13 GB

VRAM @ FP16

42 GB

Fits on (Q4)

RTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

~225 tok/s on RTX 4090 (Q4, llama.cpp)

Quantizations

GGUF Q4 · MXFP4 (native) · GGUF Q4_K_M · GGUF Q8_0

gpt-oss family

Elo progression across releases

gpt-oss-120b (High)Aug 20251845.2 gpt-oss-120b (Low)Aug 20251330.6 gpt-oss-120b (Medium)Aug 20251555.0 gpt-oss-20b (High)Aug 20251540.0 gpt-oss-20b (Low)Aug 20251071.0 gpt-oss-20b (Medium)Aug 20251367.6

Resources

HFopenai/gpt-oss-20b GHopenai/gpt-oss DOCScookbook.openai.com APIAvailable via OpenAI API

API price weights · each benchmark row carries its own source badge (see methodology)