gpt-oss-120b: 116.83B total / 5.13B active per token; gpt-oss-20b: 20.91B total / 3.61B active per token (MoE)
131,072 tokens (both variants)
Apache License 2.0; accompanying gpt-oss usage policy requires compliance with applicable law
Two original text-only, autoregressive reasoning/chat MoE checkpoints released together; the 120b and 20b names round actual totals of about 117B and 21B. The larger model has 36 layers and 128 experts per layer, the smaller 24 layers and 32 experts; both route each token to four SwiGLU experts, use a 2,880-wide residual stream, and alternate 128-token sliding-window and full attention with GQA (64 query heads, eight key/value heads) and YaRN-scaled 131,072-token context. Released expert matrices use MXFP4 (4.25 bits per parameter including scales), while attention and other non-expert components remain BF16; packed storage elements and quantization scales are not additional learned parameters. OpenAI reports that optimized MXFP4 inference fits 120b on one 80GB GPU and 20b on systems with 16GB memory, not that BF16 inference or a fully populated 128K context fits those budgets; KV cache, runtime and batch size affect memory, and CPU speed is not established by these fit claims. Trained on trillions of text tokens emphasizing STEM, code and general knowledge, with a June 2024 knowledge cutoff and harmful-content filtering; post-training uses chain-of-thought reinforcement learning similar to o3, tool-use training and deliberative safety alignment. Exact corpus composition, token budget and full post-training recipe are not disclosed in the launch report. Low/medium/high reasoning effort is selected in the system prompt, not by loading separate weights. Harmony formatting is required, separating analysis, commentary/tool recipients and final answers; supports developer-defined function calls and structured outputs. Browsing and Python require external tools and an execution harness, not capabilities that run autonomously inside the weights. In OpenAI's high-effort, quantized evaluations, AIME 2025 scores are 92.5%/91.7% without tools and 97.9%/98.7% with tools (120b/20b); HLE is 14.9%/10.9% without tools and 19.0%/17.3% with tools. SWE-bench Verified scores of 62.4%/60.7% are tool-assisted evaluation results, not standalone code-generation or demonstrated live-agent reliability. Strengths include efficient reasoning, mathematics and tool-oriented coding; factual reliability is weaker: without browsing, SimpleQA hallucination rates are 78.2%/91.4%, versus 75.0% for o4-mini in the same report. Longer reasoning increases latency and cost. OpenAI evaluated adversarially fine-tuned 120b variants for biological/chemical and cyber risk and reported that they did not reach its High-capability thresholds under the tested conditions; this is not a guarantee against downstream misuse or safety removal. Later safeguard models, community derivatives and purported unquantized conversions are outside this release grouping.