gpt-oss-120b

What gpt-oss-120b is

gpt-oss-120b is downloadable software-model weights, rather than a ChatGPT subscription. It generates text and tool requests inside an application you operate. The name describes a mixture-of-experts model with about 117 billion total parameters and 5.1 billion used per token; the smaller active number does not mean the remaining weights disappear from memory.

The downloadable checkpoint is a building block, not a complete application. A runtime loads its weights and formats messages, while your application handles files, tool access and user review. Choosing open weights changes who operates those parts; it does not automatically connect them or remove the need to check generated answers.

How it fits into a workflow

A compatible runtime loads the model and accepts messages. Your application supplies the tools, decides what they may do and returns results to the model. Choose a reasoning setting and test a small representative task before building a service around it.

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

Requirements, cost and limitations

OpenAI says the released MXFP4 model fits a single 80 GB GPU. Treat this as a publisher configuration claim, not a guarantee for arbitrary runtime, context length or simultaneous users. Harmony message formatting is required. Apache-2.0 permits use under its terms, but hardware, hosting and operations still have costs.

When to choose it and what to compare

Consider it when deployment control matters and someone can own the server. For a laptop experiment, first compare the smaller gpt-oss-20b. The August 2025 evaluation report is useful historical evidence; it does not establish a current overall winner or a price per successful task.

Access, running cost and practical requirements

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

Download weights and operate a compatible local or cloud runtime; this is not ChatGPT access.

Inputs and outputs

Text to text and tool-call messages.

License / access terms

Apache-2.0

Memory and hardware planning

117B total / 5.1B active; official MXFP4 single-80GB-GPU claim. For many ordinary PCs this points toward rented GPU or dedicated infrastructure, but local high-memory systems are not excluded.

A practical setup path

Follow the official Harmony-compatible vLLM or Transformers example. Start with one request and short context before testing simultaneous users.

A server-scale reasoning example for operators who need control over deployment and can maintain GPU infrastructure.

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 2026-10-29. Workflow suggestions and comparisons are editorial interpretation; PrisAIm has not independently benchmarked this product.