GPT OSS 20B (Greenference)

GPT OSS 20B is the smaller member of OpenAI's open-weight gpt-oss family, designed to bring reasoning and agentic capabilities into a model that organizations can run and adapt themselves. The official model card groups it alongside its larger sibling and devotes a dedicated post-training section to reasoning, variable-effort reasoning training, and agentic tool use, all wired through the Harmony chat format. The intent is a model that can step through problems deliberately, call tools when needed, and be deployed outside of a closed API, making it a practical choice for teams building reasoning assistants or agent pipelines who want full control over weights and serving.

Under the hood, GPT OSS 20B uses a Mixture-of-Experts design with roughly 20.9B total parameters but only about 3.61B active per token, which is central to its efficiency story. An independent single-GPU H100 analysis in bf16 found that, at a 2,048-token context with 64-token decode, the model delivered higher decode throughput and better tokens-per-Joule than dense baselines such as Qwen3-32B and Yi-34B, while using substantially less peak VRAM; the trade-off is somewhat higher time-to-first-token because of MoE routing. That combination of sparse activation, open weights, and tool-aware post-training makes GPT OSS 20B well suited to latency-sensitive, cost-aware deployments where on-prem or self-hosted inference is preferred.