GPT OSS 20B

OpenAI released gpt-oss-20b on August 5, 2025 as an open-weight reasoning model available under the Apache 2.0 license, alongside the larger gpt-oss-120b. The 20B variant targets edge and low-latency use cases, delivering results comparable to OpenAI o3-mini on common benchmarks while running on devices with just 16 GB of memory. The model supports adjustable reasoning effort, full chain-of-thought outputs, structured outputs, and tool use including web search and Python execution, with compatibility for the Responses API. It was trained using reinforcement learning and techniques informed by OpenAI's frontier models including o3, and strong tool-use results were observed on the Tau-Bench agentic suite and HealthBench evaluations.

OpenAI's developer documentation describes gpt-oss-20b as a medium-sized open-weight model for low-latency, local, or specialized use cases, with 21B total parameters and 3.6B active parameters. The model is offered under the Apache 2.0 license with download available via Hugging Face and supports a 131,072-token context window with 131,072 max output tokens. Key features include configurable reasoning effort across low, medium, and high levels, full chain-of-thought access for debugging and trust, parameter fine-tuning support, and native agentic capabilities covering function calling, web browsing, Python code execution, and structured outputs. Supported endpoints include Chat Completions, Responses, and Realtime translation and transcription services.

An independent arXiv evaluation (2508.12461) benchmarked gpt-oss-20b and gpt-oss-120b against six contemporary open-source LLMs from 14.7B to 235B parameters across ten benchmarks covering general knowledge, mathematics, code generation, multilingual understanding, and conversation. All models were tested unquantized under standardized inference with statistical validation. The study found gpt-oss-20B consistently outperformed gpt-oss-120B on several benchmarks such as HumanEval and MMLU despite substantially lower memory and energy per response. Both variants showed mid-tier overall performance in the open-source landscape, with relative strength in code generation but notable weaknesses on multilingual tasks, suggesting sparse scaling may not yield proportional gains.

The arXiv-hosted version of OpenAI's gpt-oss model card (2508.10925) presents gpt-oss-20b as an Apache 2.0 open-weight reasoning model with Responses API compatibility and full chain-of-thought. It is designed for agentic use cases with adjustable reasoning effort, tool use including web search and code execution, and structured outputs. The paper details OpenAI's Preparedness Framework safety testing, including adversarial fine-tuning of gpt-oss-120b to assess risk in Biological and Chemical and Cyber domains. The Safety Advisory Group concluded that even with robust fine-tuning leveraging OpenAI's training stack, the model did not reach High capability thresholds in those Tracked Categories.

OpenAI's Deployment Safety Hub mirrors the gpt-oss model card, documenting gpt-oss-20b as an Apache 2.0 open-weight text model built for agentic workflows with tool use and adjustable reasoning. The page reaffirms compatibility with the Responses API, customizable weights, full chain-of-thought visibility, and structured output support. Safety analysis describes Preparedness Framework evaluations concluding that gpt-oss models did not reach High capability in Biological and Chemical, Cyber, or AI Self-Improvement categories, even under adversarial fine-tuning reviewed by OpenAI's Safety Advisory Group. Deployers are cautioned that open weights shift safety responsibility onto downstream stakeholders.

OpenAI's gpt-oss-20b model card describes a text-only open-weight reasoning model released under Apache 2.0, developed with open-source community feedback. The document emphasizes the model's agentic workflow capabilities, including instruction following, web search, Python execution, adjustable reasoning effort, full chain-of-thought, and Structured Outputs support. Safety considerations highlight that because weights are released, determined attackers could fine-tune gpt-oss-20b to bypass safety refusals, requiring downstream developers to implement additional safeguards. OpenAI conducted scalable capability evaluations and adversarial fine-tuning testing under its Preparedness Framework, finding the default models did not reach High capability thresholds in Tracked Categories.