OpenAI: gpt-oss-20b

OpenAI: gpt-oss-20b is an open-weight language model released by OpenAI under the Apache 2.0 license. It has 21 billion total parameters and uses a Mixture-of-Experts (MoE) architecture with 3.6 billion active parameters per forward pass. The model was published by the provider on 2025-08-05. Because it is open-weight and Apache 2.0 licensed, the weights are available for inspection and self-hosting, and it sits in the text category of OpenAI's model line.

The model is text in, text out. It accepts text prompts and returns text completions, with a context window of 131,072 tokens and a maximum output of 32,768 tokens. Its knowledge cutoff is 2024-06-30, so anything after that date is outside its training data. It supports reasoning output, which you can control through the reasoning, reasoning_effort, and include_reasoning parameters when you want the model to expose or suppress its intermediate thinking.

Those capabilities point to a few practical uses. The long context suits work over large documents, such as summarizing or extracting from long reports, code files, or transcripts that would otherwise need chunking. Reasoning support makes it a reasonable fit for tasks that benefit from step by step problem solving, including math, logic, and multi step code generation or debugging. Function calling through the tools and tool_choice parameters lets you build agents that call external APIs, and structured_outputs and response_format help when downstream code needs JSON or a fixed schema rather than free text. Sampling controls such as temperature, top_p, top_k, min_p, frequency_penalty, presence_penalty, and repetition_penalty let you tune determinism and repetition for both creative and analytical workloads, while seed supports reproducible sampling.

You call the model through api14's OpenAI-compatible API using the model id "openai/gpt-oss-20b". Because the endpoint follows the OpenAI schema, existing OpenAI client libraries and request shapes work without changes. Alongside the usual fields, the model supports frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, reasoning_effort, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, and top_p, so you can control sampling, logging, reasoning, and tool use from the request itself.

The main constraint to keep in mind is modality: gpt-oss-20b only handles text, so image, audio, and other non-text inputs need a different model. The 131,072 token context and 32,768 token output ceiling are large but finite, so very long inputs must be trimmed or split. If your task depends on knowledge after 2024-06-30, the model will not know about it. For workloads where a smaller, faster, or non-text capable model fits better, api14 lists other models you can switch between without changing your integration.