OpenAI has unveiled gpt-oss-120b and gpt-oss-20b, its first fully open-weight language models since GPT-2, available under the Apache 2.0 license. The gpt-oss-120b model operates with 5.1 billion parameters per token via a mixture-of-experts architecture and performs well on core reasoning benchmarks. The gpt-oss-20b model, utilizing 3.6 billion parameters, is suitable for consumer-grade hardware with 16 GB memory. Both models feature advanced reasoning capabilities and can be tuned for speed and accuracy. Fine-tuning for risk showed these models did not reach high-risk levels, using input from expert reviewers to guide their release.
OpenAI's gpt-oss-120b model operates with 5.1 billion parameters per token via a mixture-of-experts architecture, matching or surpassing proprietary benchmarks on reasoning tasks.
The gpt-oss-20b model utilizes 3.6 billion of its parameters and is designed to run on consumer-grade hardware with 16 GB memory, enabling on-device inference.
Both gpt-oss models support advanced use cases like chain-of-thought reasoning and can be configured for varying levels of reasoning effort, balancing speed and accuracy.
OpenAI conducted adversarial fine-tuning for risk assessment, revealing the models did not achieve high-risk capability levels according to their Preparedness Framework.
Collection
[
|
...
]