gpt-oss-20b

What gpt-oss-20b is

gpt-oss-20b is a downloadable text model from OpenAI. To use it, you need a runtime: software that loads the model and runs it. This makes it a building block for a local application or a service you operate, rather than a complete assistant with all its tools already connected.

The model offers low, medium and high reasoning-effort settings, and its documentation includes local and server examples. A function-calling prototype is one possible project: the model requests an action that your application provides. Your application must connect those tools and decide what they are allowed to do.

How it fits into a workflow

Start with a small application that accepts a question and returns a text answer. If you add a tool, use one harmless operation, such as reading a fixed sample record, and inspect the requested arguments before execution. This separates whether the model proposes the right action from whether your application carries it out correctly.

Examples described as illustrative or proposed trials are editorial suggestions. Documented real builds, when available, are linked separately below.

Requirements, cost and limitations

Check that the runtime handles Harmony, the response format this model requires. Memory depends on the runtime, model representation and conversation context; “20b” is not a RAM specification. Read the model card and repository setup examples before choosing a download or hardware configuration.

When to choose it and what to compare

Consider it when downloadable weights and control of the application are important. Ollama and vLLM are possible runtime choices to investigate, not competing models. A hosted API can reduce model-operation work, while a local setup gives you responsibilities for compatibility, performance and the complete data path.

Access, running cost and practical requirements

Versioned open-weight baseline; no current frontier-rank claim. How we use this label.

Inputs and outputs

Text to text, including tool-call messages; tools are supplied by the application.

License / access terms

Apache-2.0

Memory and hardware planning

OpenAI states the MXFP4 model can run within 16 GB memory. This is a publisher fit statement, not a guarantee for every runtime, context or workload. 21B total / 3.6B active parameters does not mean only active weights need storage.

A practical setup path

Use a Harmony-compatible runtime; Transformers' chat template handles the required format. Choose low, medium or high reasoning effort, then validate output and tool permissions.

Local reasoning or tool-use prototypes when deployment control is important. More reasoning can consume more time and context; it is not a ready-connected assistant.

Running downloaded weights on your own computer has no hosted-model token bill, but hardware, power and maintenance still cost money. Rented GPUs and managed APIs charge differently. Check the model license and the runtime before spending on a deployment.

Benchmark evidence and what it means

These are selected source-reported observations, not a league table. Task, version, tool access, reasoning budget and test harness all affect a result. A historical score does not establish current leadership or accuracy on your own work.

Sources and evidence

Product facts are based on the following published sources, checked 2026-10-27. Workflow suggestions and comparisons are editorial interpretation; PrisAIm has not independently benchmarked this product.