The model architecture is extensively documented in the official model card and technical reports. It is a Mixture-of-Experts (MoE) Transformer with 24 layers, utilizing 32 experts with 4 active per token. Specific technical optimizations are disclosed, including SwiGLU activation functions (with clamping and residual connections), Grouped-Query Attention (GQA) with 8 key-value heads, and Rotary Positional Embeddings (RoPE). The model supports a native context window of 128k tokens. While the training methodology is described as a mix of reinforcement learning and advanced pre-training, the exact 'from scratch' vs. 'fine-tuned' lineage of the base weights is slightly obscured by references to 'internal frontier systems,' preventing a perfect score.

Transparency regarding the training data is minimal. Official documentation states the model was trained on a 'mostly English, text-only dataset' with a focus on STEM, coding, and general knowledge. However, there is no public breakdown of data sources, no specific proportions (e.g., web vs. books vs. code), and no detailed disclosure of filtering or cleaning methodologies. Third-party reports explicitly list training data collection and labeling as 'undisclosed.'

The model uses the 'o200k_harmony' tokenizer, which is a BPE-style tokenizer with a vocabulary size of approximately 200,000 tokens. This tokenizer is publicly available via the 'tiktoken' library and is documented as a superset of the tokenizer used in GPT-4o. It is well-integrated into the 'openai-harmony' package and reference implementations, allowing for full verification of tokenization behavior and language support alignment.

OpenAI provides precise figures for both total and active parameters. The model has 21.1 billion total parameters, with 3.6 billion active parameters per token. The MoE structure is clearly defined (32 experts, top-4 routing). Additionally, the use of MXFP4 quantization for MoE weights is explicitly documented, including how weights are packed and scaled, which provides high transparency into the model's density and memory efficiency.

There is almost no verifiable information regarding the compute resources used for training. While the hardware requirements for inference are well-documented, the training duration, GPU/TPU hours, hardware specifications used during training, and the resulting carbon footprint are entirely absent from public documentation. Claims of 'advanced pre-training' serve as marketing language rather than technical disclosure.

While OpenAI provides results for several benchmarks (MMLU, HumanEval, Tau-Bench, HealthBench), the evaluation methodology lacks full transparency. Evaluation code is not fully public, and exact prompts or few-shot examples used for official scores are not consistently disclosed. Third-party audits (e.g., arXiv:2508.17525) have noted discrepancies where the 20B model outperforms the 120B variant, suggesting non-monotonic scaling or evaluation inconsistencies that are not addressed in official docs. A -2 penalty was applied due to industry-wide concerns regarding training data overlap with common coding benchmarks like SWE-bench.

The model demonstrates strong identity consistency, correctly identifying itself as part of the GPT-OSS family in standard deployments. It is designed to use the 'Harmony' response format, which includes specific roles (System, Developer, User, Assistant, Tool) that help maintain its persona and operational boundaries. There are no documented cases of the model claiming to be a competitor's product or denying its AI nature.

The model is released under the Apache 2.0 license, which is a standard, highly permissive open-source license. This allows for commercial use, modification, and distribution without the 'copyleft' restrictions found in other licenses. The terms are clear, publicly accessible on GitHub and Hugging Face, and do not conflict with the model's stated 'open-weight' status.

Hardware requirements are exceptionally well-documented. OpenAI and third-party partners (NVIDIA, Unsloth) provide specific VRAM targets: 16GB for the 20B model using native MXFP4 quantization. Performance metrics for different hardware (Mac, H100, consumer GPUs) and token-per-second estimates are widely available. The impact of the 'reasoning effort' setting on latency is also clearly explained.

The model supports 'snapshots' to lock in specific versions, and a basic versioning system is in place. However, there is no comprehensive public changelog or detailed history of weight updates since the initial release. While the 'openai-harmony' package is versioned, the underlying model weights lack the granular semantic versioning required for a higher score, making it difficult to track silent drift or minor optimizations.