OpenAI releases gpt-oss, its first language models since GPT-2

On August 5, 2025, OpenAI released the weights of gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license. The original source supports the documentary core of the event; downloadable weights do not mean fully open data, training or service.

The decision matters because it contrasts with the strategy that made OpenAI one of the industry’s central companies: GPT-3, GPT-4, the o family and ChatGPT models have been distributed as closed services. With gpt-oss, developers, companies and government agencies can run the model on their own infrastructure, tune its behavior with specialized data and avoid having to send every query to an external API.

Two sizes for local deployment

The models are distributed under the Apache 2.0 license, a permissive license that allows commercial use, modification and redistribution. That does not mean they are completely transparent software—OpenAI has not published the training data or the full development process—but it does provide the trained parameters, the essential component needed to run and adapt the model.

The larger model, gpt-oss-120b, has 117 billion parameters in total and is designed to run on a single GPU with 80 GB of memory. The smaller gpt-oss-20b has 21 billion parameters and can run on devices with 16 GB of memory, bringing it within reach of powerful personal computers and more affordable local deployments. Document supporting the figure.

The key is their mixture-of-experts architecture, or MoE. Rather than activating every parameter for each word or text fragment, the model selects only some of its internal components. The larger model activates 5.1 billion parameters per token, while the smaller one activates 3.6 billion. This design lowers inference costs—the work required to generate a response—without sacrificing much greater total capacity. Document supporting the figure.

Both models support context windows of up to 128,000 tokens, enough to process lengthy documents, code repositories or long conversations. Document supporting the figure.

Reasoning, tools and comparative results

OpenAI places gpt-oss-120b close to o4-mini in its reasoning tests and says gpt-oss-20b achieves results similar to o3-mini on standard evaluations. The company highlights tests covering mathematics, competitive programming, general knowledge, health and tool use.

Both models can adjust reasoning effort across low, medium and high levels. In practice, developers can request a quick answer for a simple task or let the system spend more computing time on a complex problem. They are also designed to call external functions and tools, such as web searches or code execution—a relevant capability for agents that do more than simply generate text.

The comparisons should be read with caution. The results come from evaluations published by OpenAI, and benchmarks measure narrowly defined tasks rather than every enterprise use case. A model that performs well on AIME, Codeforces or HealthBench is not automatically validated for diagnosing patients, making financial decisions or automating processes without supervision.

OpenAI explicitly warns that the models do not replace health care professionals and are not intended to diagnose or treat diseases.

Open weights do not mean zero risk

Openness makes audits and adaptations easier, but it also reduces the publisher’s control over how the model is used later. Once the weights have been downloaded, an organization can run them without relying on OpenAI’s infrastructure and create versions tailored to specific fields.

The company says it subjected gpt-oss-120b to safety evaluations, including an adversarially fine-tuned version tested under its Preparedness Framework, the framework it uses to assess risks from advanced capabilities. It has also published a model card and a technical paper on those tests.

Another unusual feature is access to the model’s chain of thought. OpenAI says it has not applied direct supervision to that internal chain, with the aim of allowing researchers to study signs of deception, misuse or failures during reasoning. The company nevertheless recommends not showing it directly to end users: it may contain errors, harmful content or instructions that should not appear in the visible response.

A new option for companies and developers

The announcement does not replace OpenAI’s closed offering. The company itself presents gpt-oss as an alternative for those who prioritize cost, latency, customization or data control, while its proprietary models remain available through the API.

For a company with strict privacy requirements, hosting a model within its own network may matter more than accessing the most powerful system available. For independent developers, the 20-billion-parameter model opens the door to experimentation without a variable bill for every query. And for the open-model ecosystem, OpenAI’s entry increases competition with offerings such as Llama, Qwen, Mistral and DeepSeek. Document supporting the figure.

The decisive test will come outside the benchmarks: real deployment costs, quality in languages other than English, ease of fine-tuning and behavior in specialized applications. But the release breaks six years of a clear divide between the lab’s open research and its commercial language models.

Turning the headline into a check

The first step is to freeze the system's identity. OpenAI released the weights of gpt-oss-120b and gpt-oss-20b under the Apache 2.0 license. A commercial name may cover different revisions, automatic routes and tools. A test record should preserve date, access mode, configuration, permissions and the full output. Without that snapshot, an improvement or failure observed today cannot rigorously be attributed to the version another person will use tomorrow.

Next, turn how to audit license, memory, execution, quality and controls before hosting a model into cases with acceptance criteria. Build a local sample containing easy, ambiguous, long and deliberately impossible tasks. Record the input, the information the model may consult and what outcome would count as sufficient. A vendor-selected demonstration shows possibility; a test set preserved by the user measures reliability.

Autonomy needs a permission ladder. Reading and proposing are not the same as editing, sending or buying. A safer setup begins with read-only access, requires a preview and reserves execution for explicit approval. It also keeps a log and a rollback path. Judge the model by the errors the surrounding system contains, not by the confidence of its plan.

What the record must preserve

Cost and quality must be measured together. A cheap answer that must be reviewed from scratch can cost more than a slower but verifiable one. Measurement includes waiting, retries, consumption, human oversight and the consequences of failure. downloadable weights do not mean fully open data, training or service. That boundary turns the announcement into a testable hypothesis rather than a promise to be believed.

An evidence sheet separates four columns: what the source claims, what it shows, what it did not measure and what would change the conclusion. That discipline prevents an absence from becoming a promise and a condition from vanishing in summary. It also lets the story be updated without rewriting history from a later outcome.

Include a negative case before deciding. Find a situation where the system, rule, transaction or study does not meet the need and record the signal that would require stopping. Selected successes show that something can happen; the negative case reveals the boundary and lowers the cost of discovering it after deployment.

The skill that outlasts the announcement

A valid comparison preserves denominator and axis. It does not pit a point figure against an average, future capacity against installed capacity or a forecast against an observation. When two sources use similar language, reconstruct what they counted and over what period. If those differ, publish them as different measures instead of inventing a ranking.

The record should survive a version change. Keep URL, consultation date, document, configuration and decision. When new evidence appears, add it with its date and explain what it changes. That traceability prevents opposite errors: keeping an expired conclusion or pretending later information was known on the event date.

The transferable skill in this story is how to audit license, memory, execution, quality and controls before hosting a model. The procedure is short: name the document, preserve the date, fix the axis, find the condition and design a check that can fail. With those steps, a reader need not accept or reject the announcement by intuition; the decision follows a visible chain of evidence.