GPT OSS 20B (OVHcloud)

GPT OSS 20B is the smaller member of OpenAI's gpt-oss family of open-weight models, positioned for lower-latency inference, on-device and specialized workloads where the larger 117B sibling would be excessive. It is built as a Mixture-of-Experts transformer that shares its structural design with the 120B variant but is tuned with different hyperparameters for efficiency. Independent measurements place it at roughly 20.9B total parameters with about 3.6B active per token, using SwiGLU activations and learned attention sinks, and it is released under the permissive Apache 2.0 license that allows commercial and customized use.

In practical terms, gpt-oss-20b behaves as a text-only reasoning model aimed at chain-of-thought processing, adjustable reasoning effort, instruction following, and tool use, making it a fit for agentic workflows and developer-facing applications. A third-party deployment study on a single H100 GPU in bf16 reports higher decode throughput and better tokens-per-Joule than dense baselines such as Qwen3-32B and Yi-34B at a 2,048-token context with 64-token decode, while using substantially less peak VRAM, with the trade-off of higher time-to-first-token due to MoE routing. Those characteristics, combined with its open weights, make it well suited to private or on-premises deployments where data control and predictable cost-per-token matter more than maximizing single-prompt accuracy.