In August 2025, OpenAI released gpt-oss, a pair of open-weight language models published under the permissive Apache 2.0 license on Hugging Face. It was OpenAI’s first release of downloadable model weights since GPT-2 in 2019, a notable reversal for a company whose name had become synonymous with closed, API-only frontier models. OpenAI described them as “open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases.”
The two models used a mixture-of-experts design to run efficiently on accessible hardware. gpt-oss-120b has 117 billion total parameters with 5.1 billion active per token and fits on a single 80GB GPU through MXFP4 quantization, while gpt-oss-20b has 21 billion total parameters with 3.6 billion active and runs within 16GB of memory. Both expose full chain-of-thought reasoning, support function calling and agentic use, and offer configurable reasoning effort (low, medium, high).
The gpt-oss release mattered because it brought OpenAI back into the open ecosystem amid intense competition from open-weight families like Llama, Qwen, and DeepSeek. By shipping reasoning-capable models that run on a single GPU or a laptop-class machine, it lowered the barrier to deploying capable AI without an API dependency.