GPT-OSS 120B

GPT-OSS 120B is an OpenAI-built language model designed around explicit reasoning behavior and released as an open-weight checkpoint. Its architecture is a Mixture-of-Experts design with 117 billion total parameters and roughly 5.1 billion parameters activated per token, distributed across 128 experts. This sparse activation pattern lets the model draw on broad learned knowledge while keeping the compute per inference comparatively modest, which is well suited to step-by-step reasoning workloads where the model pauses to think before producing a final answer.

In practice, GPT-OSS 120B fits use cases that need stronger reasoning than a small dense model can offer without paying the full cost of a very large active network. Because weights are openly available, the model can be self-hosted, fine-tuned, or evaluated for downstream applications such as analytical question answering, code reasoning, and structured planning. The same MoE design that enables its reasoning depth also makes speculative decoding an effective serving strategy, where a smaller draft model proposes tokens and the full model verifies them, improving throughput for chat and agent-style deployments.