OpenAI GPT-oss-120b

The gpt-oss-120b model uses a mixture-of-experts architecture with 117B total parameters and 5.1B active parameters per token, enabling efficient inference on a single 80GB GPU like an NVIDIA H100 or AMD MI300X. The model was trained on OpenAI's harmony response format and is designed for agentic workflows, offering configurable reasoning effort across low, medium, and high settings. It provides full chain-of-thought visibility for debugging purposes and strong instruction-following capabilities, making it well-suited for production reasoning tasks and tool-augmented applications. The model is fine-tunable, allowing developers to customize it for specific use cases without being locked into a single deployment pattern.

This model was trained using a blend of reinforcement learning and techniques informed by OpenAI's most advanced internal systems, including o3 and other frontier models. It achieves near-parity with OpenAI o4-mini on core reasoning benchmarks while demonstrating strong performance on agentic evaluation suites like Tau-Bench and HealthBench, even outperforming proprietary models such as o1 and GPT-4o on tool use and few-shot function calling tasks. The Apache 2.0 license removes commercial restrictions, making it freely usable for experimentation, customization, and deployment. Third-party ecosystem work, such as Multiverse Computing's 50% compressed HyperNova variant, shows the model's flexibility for specialized deployment scenarios across cloud infrastructure and enterprise systems like IBM watsonx Orchestrate on Amazon Bedrock.