gpt-oss-20b

Tool-calling plus extended thinking covers the basic loop for agentic coding, though the absence of published coding benchmarks and context window limits confidence for repository-scale work.

Extended thinking supports traceable reasoning over audit material, and self-hosting limits data exposure. Context window is unspecified, which constrains use on large audit files.

Tool-calling enables driving browser actions through external tool interfaces. Performance depends on the surrounding harness, as the model has no native browsing or vision capability listed.

Extended-thinking mode is suited to stepwise root-cause reasoning over code and logs. Without published coding evaluations, it should be validated against the target codebase before adoption.

Extended-thinking support supports decomposing multi-step tasks, but no published benchmarks are available to rank it against alternatives on planning quality.

Open-weights status permits on-premises deployment, which keeps sensitive text inside the customer boundary. Redaction accuracy is not documented, so it requires evaluation on representative data.

Optical character recognition requires image input, which this model's listed capabilities do not include.

No vision capability is listed for this model, so it cannot process image inputs for visual question answering.

No audio output capability is listed, so the model cannot generate speech.

No audio input capability is listed, so the model cannot perform speech-to-text transcription.