gpt-oss is a family of open-weight reasoning language models from OpenAI, released in August 2025 in two sizes, gpt-oss-120b and gpt-oss-20b, under the permissive Apache 2.0 license. Among AI models & assistants, it stands out as the first OpenAI language model since GPT-2 whose weights anyone can download, run on their own hardware and fine-tune.
Both models use a mixture-of-experts transformer that activates only a small fraction of its parameters for each token, which lets the larger model run on a single 80 GB data center GPU and the smaller one on hardware with about 16 GB of memory. They are text-only models designed for chain-of-thought reasoning, tool use and agentic workflows rather than for image or audio input.
Architecture of the gpt-oss models
The gpt-oss models are decoder-only transformers in which each feed-forward layer is replaced by a mixture-of-experts (MoE) block. A small router network selects the top four experts for every token, so most expert weights stay idle during a given step. This design separates the total parameter count, which determines knowledge capacity, from the active parameter count, which determines compute cost per token.
The table below lists the published architectural figures for both variants.
| Property | gpt-oss-120b | gpt-oss-20b |
|---|---|---|
| Total parameters | About 117 billion | About 21 billion |
| Active parameters per token | About 5.1 billion | About 3.6 billion |
| Transformer layers | 36 | 24 |
| Experts per MoE layer | 128 | 32 |
| Context length | About 128,000 tokens | About 128,000 tokens |
| Typical hardware | One 80 GB GPU | Consumer GPU or laptop with 16 GB |
Attention layers alternate between full dense attention and locally banded sparse attention, which reduces memory use on long inputs. The models use grouped multi-query attention, rotary position embeddings and an extended context mechanism. The MoE weights are stored in MXFP4, a 4-bit floating-point format, which is the main reason the models fit on such modest hardware.
Reasoning effort, tools and the harmony format
The gpt-oss models are trained to produce an internal chain of thought before the final answer, similar to the OpenAI o-series reasoning models. Developers set a reasoning effort level of low, medium or high in the system prompt, trading answer quality against latency and token cost.
| Effort level | Typical behavior | Suitable use |
|---|---|---|
| Low | Short reasoning, fast responses | Chat, classification, simple extraction |
| Medium | Balanced reasoning length | General assistants and coding help |
| High | Long, detailed reasoning traces | Mathematics, complex code and analysis |
Prompts and outputs follow the harmony response format, a structured chat template with roles and separate channels for reasoning, tool calls and the final answer. Using the wrong template is a frequent cause of poor output in self-hosted deployments. The models support function calling, web browsing and Python execution as built-in tool patterns, which makes them suitable as the core of an agent.
OpenAI deliberately did not apply direct supervision to the chain of thought, so that the reasoning stays useful for monitoring model behavior. The consequence is that raw reasoning may contain unfiltered or inaccurate content and should not be shown to end users without review.
Comparison with other open-weight model families
gpt-oss competes with other open-weight model families that developers can host themselves. Its distinguishing traits are the permissive license, the MoE efficiency and its reasoning-first training, while several competitors offer multimodal input or a wider range of sizes.
| Model family | Developer | Notable characteristics |
|---|---|---|
| gpt-oss | OpenAI | Text-only MoE reasoning models, Apache 2.0 license |
| Llama | Meta | Broad size range, custom community license |
| Qwen | Alibaba | Dense and MoE variants, strong multilingual coverage |
| DeepSeek | DeepSeek | Very large MoE reasoning models, MIT-style licensing |
| Gemma | Small models, some with image input | |
| Mistral | Mistral AI | Dense and MoE models, mixed license terms |
On published reasoning benchmarks, gpt-oss-120b performs roughly at the level of OpenAI o4-mini, while gpt-oss-20b is broadly comparable with o3-mini. Such comparisons depend heavily on the effort level, tool access and evaluation setup, so independent testing on the target workload remains essential.
Deployment and customization options
Because the weights are public, organizations can run gpt-oss entirely on their own infrastructure, which keeps prompts and data inside their security perimeter. Common inference engines and local runtimes support the models, and many cloud providers host them as managed endpoints.
| Situation | Recommended choice | Reason |
|---|---|---|
| Strict data residency | Self-hosted gpt-oss-120b | No data leaves the controlled environment |
| Developer laptop or edge device | gpt-oss-20b with a local runtime | Fits in about 16 GB of memory |
| Domain-specific vocabulary | Fine-tuning with LoRA adapters | Adapts behavior without full retraining |
| Variable traffic, no GPU team | Managed cloud endpoint | Avoids operating inference hardware |
Fine-tuning is permitted by the license, and parameter-efficient methods such as LoRA keep the cost manageable. Teams should remember that fine-tuning can weaken safety behavior, so evaluation and guardrails belong in every customized deployment.
Limitations and risks
gpt-oss offers strong reasoning per unit of compute, but it has clear limitations that matter when choosing a model:
- Text only: the models accept no images or audio, unlike many multimodal assistants.
- Hallucination on factual recall: the smaller active parameter count limits stored knowledge, so retrieval or browsing improves factual accuracy.
- Safety is the deployer’s responsibility: once weights are downloaded, the original provider cannot patch or monitor them.
- Template sensitivity: deviations from the harmony format degrade output quality noticeably.
- No hosted product features: memory, file handling and moderation must be built by the deployer.
Open weights also mean that misuse cannot be revoked. OpenAI reported adversarial fine-tuning tests before release to estimate worst-case risk, yet each deployer still needs its own usage policy and monitoring.
Stripped to the essentials
gpt-oss is a pair of AI language models that OpenAI gave away for anyone to download. Most OpenAI models only run on OpenAI servers. These two can run on your own computer or company server, and you may change and reuse them freely.
The models think through a problem step by step before answering. The large version needs one powerful graphics card, while the small one runs on a good laptop. The trade-off is simple: you get control and privacy, but you also take responsibility for running and supervising the model.
Where it is applied
A mid-sized hospital network wants an assistant that summarizes clinical guidelines and drafts internal policy documents, but patient-related text may not leave its data center. Commercial cloud assistants are therefore ruled out by its compliance team.
The IT department deploys gpt-oss-120b on a single 80 GB GPU server inside the hospital network, connects it to a retrieval index of approved guidelines and sets the reasoning effort to medium for everyday questions and high for complex policy comparisons. Staff access it through an internal chat interface, and every answer cites the retrieved guideline passages.
The outcome is a private reasoning assistant with no per-token fees and full control over logs. A clinical informatics team reviews a weekly sample of answers, and the hospital keeps a human sign-off for any document that reaches patients.