Designed as the production-oriented member of the gpt-oss lineup, this model targets high-reasoning workloads that still fit on a single 80GB accelerator such as an NVIDIA H100 or AMD MI300X, with 117B total parameters and roughly 5.1B active at inference. It is released under the permissive Apache 2.0 license, so it can be downloaded, fine-tuned, and integrated into custom pipelines without copyleft or patent friction. To work correctly, it has to be driven through the harmony response format used in training, which keeps the developer experience consistent for chat, tool calls, and structured outputs.

In practice the model behaves like a reasoning-focused generalist rather than a narrow specialist. Reasoning effort can be dialed between low, medium, and high to trade latency for depth, and the full chain-of-thought is exposed for debugging and trust without being meant for end users. Native agentic capabilities make it well suited to function calling and multi-step tool orchestration, while parameter fine-tuning lets teams specialize it for domain tasks. A long context window supports extended document analysis and multi-turn agent runs, which combined with open weights makes it a strong fit for teams that want controllable, self-hosted reasoning without sacrificing developer ergonomics.