# GPT-OSS 20B

> OpenAI's open-weight model designed for powerful reasoning, agentic tasks, and versatile developer use cases - optimized for lower latency and specialized use-cases at the edge

**ID**: `openai/gpt-oss-20b`  
**Creator**: OpenAI  
**Category**: text  
**Context**: 131K tokens  
**Input**: $0  
**Output**: $0  
**Released**: 2025-04-08  
**Web page**: https://anyrouter.dev/model/openai/gpt-oss-20b

**Input modalities**: text  
**Output modalities**: text  

**Tokenizer**: GPT

**Capabilities**: chat, reasoning, function-calling, streaming, coding

## Usage

**Endpoint**: `POST https://anyrouter.dev/api/v1/chat/completions`

```bash
curl https://anyrouter.dev/api/v1/chat/completions \
  -H "Authorization: Bearer $ANYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "openai/gpt-oss-20b",
  "messages": [
    {
      "role": "user",
      "content": "Say hi in 3 words."
    }
  ]
}'
```

## Providers

| Provider | Input | Output |
| --- | --- | --- |
| aihubmix-byok | $0 | $0 |
| huggingface-byok | $0 | $0 |
| deepinfra-byok | $0 | $0 |
| cline-byok | $0 | $0 |
| openrouter-byok | $0 | $0 |
| ovhcloud-byok | $0 | $0 |
| router-byok | $0 | $0 |
| vercel-byok | $0 | $0 |
| pollinations | $0 | $0 |
| siliconflow-byok | $0 | $0 |
| nvidia-byok | $0 | $0 |
| nousresearch-byok | $0 | $0 |

## Benchmarks

- **GPQA Diamond**: 0.7
- **HLE**: 0.1
- **τ²-Bench Telecom**: 0.6
- **SciCode**: 0.3
- **OTIS Mock AIME**: 0.7

Benchmark data from [Epoch AI](https://epoch.ai/benchmarks) (CC-BY 4.0).

## Model card

- Architecture: mixture of experts [3]
- Total parameters: 20.9B [3,1]
- Active parameters: 3.6B [3,1]
- Native context: 131072 tokens [3,2]
- Training cutoff: 2024-06 [3]
- License: Apache-2.0, open weights (https://huggingface.co/openai/gpt-oss-20b) [1]

### Strengths [1]

- Runs within 16 GB of memory with MXFP4 weights
- Reasoning effort set to low, medium or high
- Native function calling, web browsing and Python execution

### Known limits [3,1]

- Hallucinates more than o4-mini on SimpleQA and PersonQA
- Raw chain of thought can hallucinate; filter it before showing users
- Works only with the harmony response format

### Use cases [1]

- Low-latency, local or specialized use
- Agentic tasks with function calls
- Fine-tuning on consumer hardware

### API notes [1]

- Reasoning level set in the system prompt, e.g. "Reasoning: high"
- Supports function calling and structured outputs

### Links

- announcement: https://openai.com/index/introducing-gpt-oss/
- model card: https://huggingface.co/openai/gpt-oss-20b
- paper: https://arxiv.org/abs/2508.10925

### Sources

1. [openai/gpt-oss-20b model card](https://huggingface.co/openai/gpt-oss-20b/raw/main/README.md) (accessed 2026-10-14)
2. [gpt-oss-20b config.json](https://huggingface.co/openai/gpt-oss-20b/raw/main/config.json) (accessed 2026-10-14)
3. [gpt-oss-120b & gpt-oss-20b Model Card](https://arxiv.org/abs/2508.10925) (accessed 2026-10-14)

## Source

- Upstream docs: https://developers.cloudflare.com/models/gpt-oss-20b/