The model's architecture is extensively documented as a sparse Mixture-of-Experts (MoE) transformer. Technical details include the use of 128 routed experts with a specialized shared expert layer, Grouped-Query Attention (GQA) with 96 heads, and a Multi-Token Prediction (MTP) layer for speculative decoding. The implementation is integrated into the public 'transformers' library, and the 'Slime' training framework is open-sourced. While the base model is named and its evolution from the GLM-4 series is clear, the specific pretraining methodology for the 'Air' variant versus the flagship model is only partially differentiated in public reports.

Z.ai discloses a total training corpus of approximately 23 trillion tokens, with a breakdown of 15 trillion for general domain data and 8 trillion focused on code, reasoning, and math. Methodologies such as 'SemDeDup' for semantic deduplication and 'quality-aware sampling' are mentioned in technical reports. However, specific data sources (e.g., which specific web crawls or book datasets) are not named, and the exact proportions of the final mixture are not provided, falling short of high-transparency standards for dataset disclosure.

The tokenizer is publicly available via the Hugging Face repository with a stated vocabulary size of 151,552 tokens. It is based on the tiktoken implementation and is fully integrated into standard inference frameworks like vLLM and SGLang. The vocabulary size and tokenization approach are consistent across official documentation and third-party implementation files.

The model clearly distinguishes between its total parameter count (106 billion) and its active parameters (12 billion per forward pass). It specifies the activation of 9 experts per token (8 routed + 1 shared). This level of detail for an MoE architecture is exemplary, though a full layer-by-layer parameter breakdown (e.g., attention vs. FFN weights) is not explicitly tabulated in the primary model card.

While the hardware used for fine-tuning and inference is mentioned (e.g., H100 and H200 clusters), there is no public disclosure of the total GPU/TPU hours required for the 23T token pretraining phase. Furthermore, there is no provided calculation of the carbon footprint or estimated total compute cost, which are requirements for a high score in this category.

Z.ai provides results for 12 industry-standard benchmarks and has released a specific dataset of 52 coding tasks with full agent trajectories for reproducibility. However, the evaluation code for the broader benchmark suite is not fully public, and some results rely on 'internal benchmarks' or specific frameworks (like Terminus) that require additional setup to verify. The use of 'Avg@32' for math benchmarks is noted but lacks the exact seeds/prompts for full bit-for-bit reproduction.

The model consistently identifies as part of the GLM-4.5 series and maintains a clear distinction between its 'Thinking' and 'Non-Thinking' modes. It does not exhibit identity confusion with models from other providers (e.g., GPT or Claude) in official documentation or API responses. Versioning is clear within the Z.ai ecosystem.

The model weights and associated code are released under the highly permissive MIT License, which explicitly allows for commercial use and derivative works. This is a significant improvement over previous GLM versions that used custom, more restrictive licenses. The license is clearly stated on Hugging Face, GitHub, and in official blog posts.

VRAM requirements are well-documented for various precisions (BF16, FP8) and context lengths. Documentation specifies that the Air variant can run on 2x H100 GPUs for the full 128K context or even consumer-grade hardware (like a single RTX 4090) when using 4-bit quantization. While detailed quantization-accuracy tradeoff curves are missing, the basic requirements are realistic and verified by community testing.

The model uses a versioning system (GLM-4.5-Air), but a detailed, public changelog tracking specific weight updates or 'silent' alignment shifts is not maintained. While the release date is clear, there is no formal deprecation policy or historical version access for intermediate checkpoints during the post-training phase.