benchmarks
GPT-OSS-20B is OpenAI’s open-weight, text-only reasoning model, released under the Apache 2.0 license for local deployment and customization. It supports adjustable reasoning effort and tool use, with the released quantized weights designed to run within approximately 16 GB of memory.
Benchmark results
Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.
| Benchmark | Score | Source |
|---|---|---|
| GPQA DiamondPhD-level science questions | 60.8%27 Aug 2026 | View |
| OTIS Mock AIME 2024-25Competition mathematics | 65.3%27 Aug 2026 | View |
| Terminal-BenchLong-horizon command-line agent tasks | 3.4%3 Nov 2025 | View |
Weighted mean of 3 verified benchmark results: GPQA Diamond 60.8%, OTIS Mock AIME 2024-25 65.3%, Terminal-Bench 3.4%.
Capabilities
- Supports low, medium and high reasoning effort to balance reasoning depth against latency.
- Supports function calling and agentic workflows using externally provided tools.
- Provides open weights under Apache 2.0 for customization, fine-tuning and commercial deployment.
- Supports structured outputs for integration with applications.
- Offers a 131,072-token context window for long text inputs.
- Uses a mixture-of-experts architecture and released MXFP4 quantization to support comparatively compact local deployment.
What improved
- Compared with GPT-2, adds explicitly trained reasoning with adjustable reasoning effort.
- Compared with GPT-2, adds native tool-use support for agentic applications.
- Compared with GPT-2, expands the documented context window to 131,072 tokens.
Known flaws & limitations
- Text-only model: it does not natively process images, audio or video.
- Can hallucinate factual information, so consequential answers require independent verification.
- Its approximately 16 GB memory target depends on the released quantization and a compatible inference implementation; runtime requirements vary with workload.
- Requires an external serving stack and correct handling of OpenAI’s Harmony conversation format.
Risks, safety & security issues
- Open weights allow third parties to modify the model and weaken its original safety behavior.
- Agentic deployments must defend against prompt injection and prevent untrusted content from authorizing tool actions.
- Unrestricted tool permissions can turn erroneous model decisions into consequential external actions.
- Safety findings for the original release do not establish the safety of subsequently fine-tuned or otherwise modified versions.
Good fit for
- Locally hosted text assistants where infrastructure and data handling must remain under the operator’s control.
- Coding and reasoning assistants with testing, review and constrained tool access.
- Custom domain assistants built through fine-tuning and application-specific evaluation.
- Long-document analysis with source-grounded answers and independent factual checks.
Poor fit for
- Applications requiring native image, audio or video understanding.
- Unsupervised medical, legal or financial decisions requiring reliably correct factual answers.
- Autonomous agents with unrestricted access to sensitive data or consequential actions.
- Devices without sufficient memory or a compatible inference stack.
References
Profile last reviewed 10/21/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.