GLM-4.5

The model's architecture is extensively documented in the official technical report (arXiv:2508.06471) and blog. It is a sparse Mixture-of-Experts (MoE) transformer with 355B total and 32B active parameters. Specific technical details are provided, including the use of Grouped-Query Attention (GQA) with 96 heads, sigmoid-gated routing, QK-Norm for stability, and Multi-Token Prediction (MTP) layers for speculative decoding. The 'Thinking Mode' implementation is described as a hybrid reasoning framework. While the base model is clearly defined, the exact initialization state (whether from a previous GLM version or scratch) is less explicitly detailed than the structural components.

Z.ai discloses a total training corpus of 23 trillion tokens, with a breakdown of 15T general-domain tokens and 7T tokens dedicated to code and reasoning. Documentation mentions specific cleaning and filtering techniques like SemDeDup and quality-tiered up-sampling. However, the specific sources of the 'general-domain' data remain vague (e.g., 'webpages, books, papers'), and no detailed percentage breakdown by language or specific dataset names is provided, which is a significant gap for a model of this scale.

The tokenizer is publicly accessible via the official Hugging Face repository and integrated into standard frameworks like transformers, vLLM, and SGLang. It supports a 128,000-token context window. While the vocabulary size and specific tokenization algorithm (likely a Tiktoken-based BPE variant consistent with previous GLM models) are verifiable through the code, the official technical report lacks a dedicated section detailing the tokenizer's training data alignment or specific compression efficiency metrics for non-English languages.

Transparency regarding parameter counts is exemplary. The provider clearly distinguishes between total parameters (355B) and active parameters (32B) for the MoE architecture. This distinction is maintained across all official documentation, model cards, and marketing materials, preventing the common industry practice of inflating capability claims based on total MoE size. Architectural ratios (e.g., active vs. total) are easily verifiable through the released model weights.

Information regarding training compute is almost entirely absent. While the 'slime' reinforcement learning infrastructure is mentioned, there is no disclosure of total GPU/TPU hours, specific hardware cluster configurations used for the 23T token pre-training, or the estimated carbon footprint. The only hardware information relates to downstream inference requirements rather than upstream training resources.

Z.ai provides results for 12 industry-standard benchmarks (MMLU Pro, AIME24, SWE-bench, etc.) and has released the 'CC-Bench' dataset and agent trajectories on Hugging Face to facilitate verification of its agentic claims. However, the full evaluation code for all reported benchmarks is not consolidated in a single reproducible repository, and some results rely on 'internal evaluation suites' or proprietary models (like GPT-4o) as judges, which limits independent verification.

The model demonstrates high identity consistency, correctly identifying itself as GLM-4.5 and maintaining awareness of its 'Thinking' vs. 'Non-Thinking' modes. It does not exhibit the identity confusion common in fine-tuned models that claim to be GPT-4. Versioning is clear, and the model's self-description aligns with the technical specifications provided by Z.ai.

The model weights and code are released under the MIT License, which is explicitly stated in the GitHub repository, Hugging Face model card, and official blog. This is a highly transparent, permissive license that clearly allows for both commercial use and derivative works without the 'janky' custom restrictions seen in earlier GLM releases.

Hardware requirements are thoroughly documented for various precisions (FP16, BF16, FP8). Detailed VRAM estimates are provided for both the flagship 355B model and the 'Air' variant, including specific GPU counts (e.g., 16x H100 for BF16) and system RAM requirements (1TB+). The impact of context length on memory is also addressed, though more granular data on quantization-accuracy tradeoffs would further improve this score.

Z.ai uses a clear naming convention (GLM-4.5, GLM-4.5-Air, GLM-4.5V), but a formal, public changelog tracking minor weight updates or 'silent' API optimizations is lacking. While the transition from GLM-4 to GLM-4.5 is well-documented, there is limited information on how the model's behavior is maintained or monitored for drift over time following the initial release.