All AI models
Open-weightVerified August 11, 2026

OpenAI

gpt-oss-20b

gpt-oss-20b is a serious local-first option when you want OpenAI's open-weight reasoning line without the hardware footprint of a 120B-class model. It is especially useful for controlled experiments, on-device services, and privacy-sensitive text workflows, as long as you are comfortable operating the runtime yourself.

License
Apache 2.0
Context
128K tokens
Modalities
Text
Parameters
21B total, 3.6B active per token

What the license means

OpenAI releases gpt-oss-20b under Apache 2.0, subject to the gpt-oss usage policy. It is an open-weight model, not an OpenAI API model.

Open-weight, open source, and no-cost API access mean different things. Always review the linked terms and your own data, hosting, and compliance requirements before shipping.

API and operating cost

Input / 1M
No OpenAI API price
Output / 1M
No OpenAI API price

gpt-oss is not served through the OpenAI API or ChatGPT. Any per-token cost comes from your own infrastructure or a hosting provider.

Check current provider price

Where it fits

  • Local text reasoning
  • On-device or edge services
  • Private prototypes and fine-tuning

Where it does not fit

  • Vision or audio tasks
  • People who want a managed ChatGPT or OpenAI API endpoint

Local deployment reality

OpenAI states that the native MXFP4 gpt-oss-20b release can run with 16GB of memory, which makes it one of the more approachable current reasoning models for local or edge deployment.

vLLMOllamallama.cppTransformers

Practical setup notes

  • Start with a runtime that supports the model's native format before trying community conversions.
  • Benchmark your prompt length and concurrent requests because 16GB feasibility is not a throughput guarantee.

Capability notes

Reasoning
Configurable reasoning effort through supported runtimes
Tools and output
Tool use, function calling, structured output through common runtimes

Limitations to plan around

  • Text-only model.
  • Self-managed weights still incur hardware, storage, and operations costs.

Continue comparing

Related model decisions

See all reviewed models
Open-weight

gpt-oss-120b

Private-cloud reasoning services

Open source

Qwen3.6

Local coding assistance

Open model family

Gemma 4

Right-sized open-model experiments

Editorial note: This is a dated decision page, not a benchmark leaderboard or a substitute for legal, security, or infrastructure review. Model and pricing facts can change after August 11, 2026; use the linked primary sources before committing budget or customer data.