# gpt-oss-20b

> OpenAI's 21B open-weight MoE reasoning model that runs in 16GB

- Maker: OpenAI
- Line: gpt-oss (Open Weight)
- Canonical page: https://ai-tldr.dev/models/gpt-oss-20b/
- Updated: 2026-06-18

## In short

gpt-oss-20b is the smaller of OpenAI's two open-weight language models, released August 5, 2025 alongside gpt-oss-120b. It has a 131K-token context window and costs $0.029 per million input tokens and $0.14 per million output tokens. You can use it through OpenRouter, Hugging Face and Ollama.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 131K (max output 131K) | [1] |
| Input price | $0.029 / 1M tokens | [2] |
| Output price | $0.14 / 1M tokens | [2] |
| Licence | Apache 2.0 · open weights | [1] |
| Released | 5 August 2025 | [3] |
| Parameters | 20.9B total / 3.6B active (MoE) | [1] |
| Knowledge cutoff | June 2024 | [1] |

## Sources

1. [openai/gpt-oss-20b — Hugging Face model card](https://huggingface.co/openai/gpt-oss-20b)
2. [gpt-oss-20b — OpenRouter pricing & benchmarks](https://openrouter.ai/openai/gpt-oss-20b)
3. [Introducing gpt-oss — OpenAI announcement](https://openai.com/index/introducing-gpt-oss/)

## Overview

gpt-oss-20b is the smaller of OpenAI's two open-weight language models, released August 5, 2025 alongside gpt-oss-120b. It is a 20.9B-parameter Mixture-of-Experts transformer that activates only about 3.6B parameters per token, and it ships under the permissive Apache 2.0 license — part of the first open-weight model release from OpenAI since GPT-2. The weights are freely downloadable from Hugging Face and the openai/gpt-oss GitHub repository.

The headline feature of gpt-oss-20b is that it fits in roughly 16GB of memory. Its MoE weights are quantized to MXFP4 out of the box (a 12.8 GiB checkpoint), so the model runs on a single high-end laptop or consumer GPU rather than a datacenter card. That makes it OpenAI's pick for on-device reasoning, local inference, and rapid iteration. OpenAI reports it delivers results similar to its proprietary o3-mini on common benchmarks, even edging it out on competition math and health questions.

Like its larger sibling, gpt-oss-20b exposes configurable reasoning effort (low / medium / high) with visible chain-of-thought, and natively supports function calling, web browsing, Python tool use, and structured outputs via OpenAI's harmony response format. It is text-only, has a June 2024 knowledge cutoff and a 131,072-token context window. Because it is open-weight rather than served first-party, it runs across many hosts — Ollama, LM Studio, vLLM, llama.cpp, Hugging Face, OpenRouter, Fireworks, Together, AWS, Azure and others — each setting its own price, with several offering a free tier.

## Benchmarks

- MMLU: 85.3% ([source](https://arxiv.org/html/2508.10925v1))
- GPQA Diamond (no tools): 71.5% ([source](https://arxiv.org/html/2508.10925v1))
- AIME 2024 (no tools): 92.1% ([source](https://arxiv.org/html/2508.10925v1))
- AIME 2025 (with tools): 98.7% ([source](https://arxiv.org/html/2508.10925v1))
- SWE-bench Verified: 60.7% ([source](https://arxiv.org/html/2508.10925v1))
- HealthBench: 42.5% ([source](https://arxiv.org/html/2508.10925v1))
- Humanity's Last Exam (with tools): 17.3% ([source](https://arxiv.org/html/2508.10925v1))
- Tau-Bench Retail: 54.8% ([source](https://arxiv.org/html/2508.10925v1))

## Pricing

- Input: $0.029 / 1M tokens per 1M tokens
- Output: $0.14 / 1M tokens per 1M tokens

gpt-oss-20b is open-weight, so it is not served first-party by OpenAI — prices vary by host. Figures shown are a representative low rate from OpenRouter providers; several hosts also offer a free tier (e.g. openai/gpt-oss-20b:free). Self-hosting incurs only your own compute cost.

Source: https://openrouter.ai/openai/gpt-oss-20b

## Strengths

- Runs on-device in ~16GB of memory thanks to native MXFP4 quantization (12.8 GiB checkpoint)
- Strong reasoning for its size — comparable to OpenAI o3-mini, beating it on competition math and health
- Permissive Apache 2.0 license allowing commercial use, fine-tuning, and self-hosting
- Efficient MoE design: only ~3.6B of 20.9B parameters active per token for low latency
- Configurable reasoning effort (low/medium/high) with full chain-of-thought visibility
- Native agentic tooling: function calling, web browsing, Python execution, structured outputs
- 131K-token context window for long documents and agent traces
- Fine-tunable on consumer hardware with no vendor lock-in

## Best for

- On-device and local reasoning assistants that keep data off the cloud
- Edge and offline deployments where a 16GB memory footprint is the constraint
- Cost-controlled high-volume inference via self-hosting or cheap third-party providers
- Agentic workflows needing tool calling, browsing, and code execution at low latency
- Fine-tuning an open reasoning model for domain-specific applications
- Math, science, and coding tasks where o3-mini-class quality is enough
- Rapid prototyping and research into chain-of-thought with fully visible reasoning traces

## How to access

- [OpenRouter](https://openrouter.ai/openai/gpt-oss-20b) — model id `openai/gpt-oss-20b`
- [Hugging Face](https://huggingface.co/openai/gpt-oss-20b) — model id `openai/gpt-oss-20b`
- [Ollama](https://ollama.com/library/gpt-oss) — model id `gpt-oss:20b`
- [OpenAI API](https://developers.openai.com/api/docs/models/gpt-oss-20b) — model id `gpt-oss-20b`

## gpt-oss (Open Weight) versions

- [gpt-oss-120b](https://ai-tldr.dev/models/gpt-oss-120b/) (5 August 2025)
- gpt-oss-20b (5 August 2025) — this page

## FAQ

### What hardware do I need to run gpt-oss-20b?

Because its Mixture-of-Experts weights ship in native MXFP4 quantization (a 12.8 GiB checkpoint), gpt-oss-20b runs within about 16GB of memory — small enough for a high-end laptop or a single consumer GPU. It has 20.9B total parameters but activates only about 3.6B per token, keeping latency low.

### Is gpt-oss-20b free and open source?

The weights are released under the permissive Apache 2.0 license, so you can download, run, fine-tune, and deploy gpt-oss-20b commercially without copyleft restrictions. Running it still costs compute, but several hosts (including an OpenRouter free tier) offer it at zero or near-zero cost, and self-hosting incurs only your own hardware cost.

### How does gpt-oss-20b compare to OpenAI's o3-mini?

OpenAI reports gpt-oss-20b delivers results similar to its proprietary o3-mini on common benchmarks, and matches or exceeds it on competition mathematics (AIME) and health questions (HealthBench) — while being fully open-weight, self-hostable, and able to run on-device in about 16GB.

### Is gpt-oss-20b multimodal?

No. gpt-oss-20b is text-only — it does not accept image, audio, or video input. It supports a 131,072-token context window, configurable reasoning effort (low/medium/high) with visible chain-of-thought, and native tool use including function calling, browsing, and Python execution via OpenAI's harmony response format.

## Links

- [Introducing gpt-oss — OpenAI announcement](https://openai.com/index/introducing-gpt-oss/)
- [openai/gpt-oss-20b — Hugging Face model card](https://huggingface.co/openai/gpt-oss-20b)
- [gpt-oss-120b & gpt-oss-20b Model Card (arXiv 2508.10925)](https://arxiv.org/abs/2508.10925)
- [gpt-oss-20b — OpenAI API docs](https://developers.openai.com/api/docs/models/gpt-oss-20b)
- [openai/gpt-oss — official GitHub repository](https://github.com/openai/gpt-oss)
- [gpt-oss playground (gpt-oss.com)](https://gpt-oss.com/)
- [gpt-oss-20b — OpenRouter pricing & benchmarks](https://openrouter.ai/openai/gpt-oss-20b)
- [gpt-oss-20b benchmark scores (arXiv model card HTML)](https://arxiv.org/html/2508.10925v1)
- [gpt-oss:20b — Ollama library](https://ollama.com/library/gpt-oss)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/gpt-oss-20b/