OpenAI’s official GPT-OSS model card describes GPT-OSS-20B and GPT-OSS-120B as open-weight, text-only reasoning models, and outlines shared capabilities and safety considerations. OpenAI’s GPT-OSS-20B model page specifies 21 billion total parameters and 3.6 billion active parameters, and positions the model for low-latency, local, or specialized use cases.
One important API distinction: Responses API compatibility does not mean OpenAI hosts GPT-OSS-20B through the OpenAI API. OpenAI’s Help Center says the open-weight models are not served through that API or available in ChatGPT. For a hosted option, Microsoft Foundry lists GPT-OSS-20B as generally available.
What the official model card says
OpenAI’s model card introduces GPT-OSS-20B alongside GPT-OSS-120B. Its overview attributes these capabilities to the model family:
- Text-only reasoning and instruction following.
- Tool use, including web search and Python code execution, for agentic workflows.
- Adjustable reasoning effort, customization, and Structured Outputs.
- Compatibility with the Responses API.
These are shared overview statements, not a complete list of features independently demonstrated for every deployment or provider. The scalable safety-capability evaluations and adversarial fine-tuning findings summarized in the Deployment Safety Hub section concern GPT-OSS-120B; do not transfer those findings to GPT-OSS-20B. This section-specific finding does not establish whether the complete model card contains separate GPT-OSS-20B evaluations.
OpenAI’s GPT-OSS-20B model page gives its specific parameter profile: 21B total parameters and 3.6B active parameters. Microsoft Foundry’s catalog lists a 131.072k-token context window and token limit for its GPT-OSS-20B listing. OpenAI describes the model as intended for low-latency, local, or specialized use cases. (OpenAI model page; Microsoft Foundry catalog)
The Hugging Face model card says the models were trained on the Harmony response format and should only be used with that format to work correctly. It also says the MoE weights use MXFP4 quantization and that GPT-OSS-20B can run within 16 GB of memory. Treat that memory figure as the card’s stated guidance, not a guarantee of a particular speed or production configuration.
Is GPT-OSS-20B a fit for your workload?
Consider it when you need an open-weight model for local or hosted deployment, customization, or specialized reasoning workflows—and when your own tests show that its answer quality and operating characteristics meet your requirements. OpenAI’s overview describes reasoning and tool-use capabilities, but does not establish a universal performance ranking for GPT-OSS-20B across workloads.
A Hugging Face discussion participant reports a SimpleQA score of 6.7/100 and describes errors in a demo response about a television show. These are community-reported claims and an anecdote, not an independently verified benchmark or an official conclusion. Treat them as a reason to test your own tasks, not as a definitive verdict on every use case.
A practical evaluation is to assemble representative prompts from your actual workload and score the outputs against known answers or acceptance criteria. For example, if your application answers questions over internal documents, test whether it retrieves the right passages, answers faithfully from them, and abstains when the evidence is missing. Track accuracy, unsupported claims, tool-call success, latency, and cost under the provider or hardware configuration you expect to use. This is a proposed evaluation method—not a claim that GPT-OSS-20B has achieved any particular result on those measures.
Can you use it through a production API?
Not through the OpenAI API, according to OpenAI’s Help Center. Although the model card says the models are compatible with the Responses API, OpenAI separately states that GPT-OSS models are not served through the OpenAI API or available in ChatGPT. Compatibility does not establish that OpenAI hosts the model for API customers. (OpenAI Help Center; model card)
Microsoft Foundry lists GPT-OSS-20B as Generally Available and provides a catalog entry with context and token-limit information. The catalog directs readers to view pricing; check the current pricing for your chosen deployment before estimating spend. (Microsoft Foundry catalog)
Before choosing a hosted API, confirm the provider’s current model identifier, supported endpoint and features, regional availability, usage limits, and pricing. Compare those terms with self-hosting if control over infrastructure, data location, or model customization is important. The sources cited here do not establish that TokenForge offers GPT-OSS-20B.
License and safety considerations
The model card names the Apache 2.0 license and OpenAI’s gpt-oss usage policy. OpenAI warns that open-weight models have a different risk profile from provider-hosted proprietary models: after weights are released, others may fine-tune them to bypass refusals or optimize for harm, and OpenAI cannot add further mitigations to or revoke access to those released weights. The card says developers and enterprises may need additional safeguards in some contexts; it does not prescribe one universal control checklist. Review the complete license, usage policy, and provider terms for your application. (OpenAI model card)
Bottom line
GPT-OSS-20B is an open-weight, text-only reasoning model with 21B total parameters and 3.6B active parameters, positioned by OpenAI for lower-latency, local, or specialized uses. The model card describes shared family capabilities. The scalable safety-capability evaluations and adversarial fine-tuning findings summarized in the Deployment Safety Hub section concern GPT-OSS-120B; that section-specific scope does not establish whether the complete model card contains separate GPT-OSS-20B evaluation results. For hosted use, distinguish API compatibility from provider availability: OpenAI says it does not serve the model through the OpenAI API, while Microsoft Foundry lists it as generally available. Make the final choice using workload-specific tests and the selected provider’s current terms.
Sources
GPT-OSS-120B & GPT-OSS-20B Model Card — OpenAI
gpt-oss-120b & gpt-oss-20b Model Card — OpenAI Deployment Safety Hub
openai/gpt-oss-20b — Hugging Face model card
OpenAI open-weight models (gpt-oss) — OpenAI Help Center