OpenAI Model Card Analysis

The gpt-oss models are autoregressive Mixture-of-Experts (MoE) transformers, building on GPT-2 and GPT-3. gpt-oss-120b has 116.8B total and 5.1B active parameters, while gpt-oss-20b has 20.9B total and 3.6B active parameters. Quantization to MXFP4 significantly reduces memory footprint, enabling deployment on a single 80GB GPU (120b) or 16GB memory (20b). Both models use a 2880 residual stream dimension, root mean square normalization, and Pre-LN placement. MoE blocks utilize gated SwiGLU activation with 128 experts for 120b and 32 for 20b, selecting top-4 experts per token. Attention blocks alternate between banded window and fully dense patterns, using Grouped Query Attention (GQA) with rotary position embeddings and YaRN for extended context. Training data focused on STEM, coding, and general knowledge, filtered for harmful content. Post-training involved CoT RL for reasoning and tool use, leveraging a "harmony chat format" for agentic features.

The gpt-oss models demonstrate strong capabilities across various benchmarks. On canonical reasoning tasks, gpt-oss-120b rivals OpenAI 04-mini, particularly in math (AIME 2024: 96.6% with tools) due to effective use of long CoTs. It also performs well on college-level exams (MMLU: 90.0%). For agentic tasks, gpt-oss-120b excels in coding (Codeforces Elo: 2622 with tools) and software engineering (SWE-Bench Verified: 62.4%). The models support "low," "medium," and "high" reasoning levels, showing smooth test-time scaling of accuracy with increased CoT length. Agentic tool use includes browsing, Python, and custom developer functions. In health, gpt-oss-120b nearly matches OpenAI 03 on HealthBench and outperforms GPT-4o, OpenAI 01, and 04-mini on several metrics, making it impactful for global health. Multilingually, gpt-oss-120b approaches 04-mini performance across 14 languages.

Safety is paramount. GPT-OSS models are trained with deliberative alignment for content refusal, jailbreak robustness, and instruction hierarchy adherence. Standard disallowed content evaluations show performance on par with or outperforming OpenAI 04-mini on production benchmarks, despite some underperformance in instruction hierarchy adherence. Hallucinated chains of thought are a known characteristic, and OpenAI has opted not to directly optimize CoT to allow for developer-implemented monitoring systems. While gpt-oss-120b shows some underperformance in hallucination rates compared to 04-mini on SimpleQA and PersonQA, this is expected for smaller models and can be mitigated by tool use. Fairness and Bias evaluations (BBQ) show parity with OpenAI 04-mini.