gpt-oss:20b

gpt-oss:20b is the smaller sibling in OpenAI's open-weight gpt-oss series, designed for powerful reasoning, agentic tasks, and versatile developer workflows. It uses a mixture-of-experts layout with roughly 21 billion total parameters but only about 3.6 billion active per token, which keeps latency low and makes the model practical for local or specialized deployments compared with its 120B counterpart. The release emphasizes configurable reasoning effort (low, medium, high), full chain-of-thought visibility for debugging, and native function-calling support, making it well suited for agentic pipelines where step-by-step reasoning needs to be inspected or steered.

On Ollama, the model is published under the gptoss architecture tag with about 20.9B parameters and MXFP4 quantization, shipping as a roughly 14GB artifact under the permissive Apache 2.0 license. The Hugging Face model card explicitly describes gpt-oss-20b as targeting "lower latency, and local or specialized use cases" rather than the high-throughput production scenarios aimed at the 120B variant. Ollama's cloud mode lets the same weights run on hosted infrastructure without requiring a powerful local GPU, which broadens practical fit for developers who want the same behavior across desktop and server environments. Together, these properties position gpt-oss:20b as an accessible open model for reasoning-heavy experimentation, fine-tuning, and tool-using agents.