GLM-4.7

The model is explicitly identified as a Mixture of Experts (MoE) architecture with 358B total parameters, building upon the GLM-4 series. Documentation describes high-level features like 'Interleaved Thinking' and 'Preserved Thinking' for agentic workflows. However, while it is described as a 'General Language Model' (GLM) architecture, specific technical details regarding the internal layer configurations, routing mechanisms, or the exact pre-training methodology are not provided in a formal technical paper, relying instead on blog posts and model cards.

There is no public disclosure of the specific training data sources, dataset proportions, or filtering methodologies. Official documentation mentions 'large-scale' training but lacks any breakdown (e.g., code vs. web vs. books). The data collection process remains opaque, and no sample data or detailed composition statistics are available to the public, falling into the category of 'proprietary dataset' claims.

The tokenizer is publicly available via the Hugging Face repository and is compatible with standard libraries like 'transformers'. Vocabulary size and tokenization behavior are verifiable through the provided code snippets and API documentation. It supports multilingual inputs (English and Chinese) with documented token-to-word ratios (~0.75 for English, 1.5 for Chinese), though detailed training alignment for the tokenizer itself is not fully explored in the documentation.

The total parameter count is clearly stated as 358B. However, for the flagship GLM-4.7 variant, the number of active parameters per token is not explicitly disclosed in the primary documentation, unlike the 'Flash' variant which specifies 3B active out of 30B. This lack of clarity on active parameters for the main model makes it difficult to assess true computational density and efficiency.

Information regarding training compute is extremely limited. While community discussions (AMA) mention the use of resources equivalent to roughly 2,000 H100/H800 GPUs for post-training experiments, there is no official disclosure of total GPU hours, hardware specifications for the full pre-training, carbon footprint, or estimated training costs in any formal documentation.

While the model provides scores for numerous public benchmarks (SWE-bench, MMLU-Pro, HLE), the evaluation code and exact prompts used for these specific results are not fully public. Users on Hugging Face have raised questions about replicating 'search-agent' and 'context management' benchmarks (BrowseComp/HLE), noting that the internal frameworks used for these evaluations have not been open-sourced, hindering independent verification.

The model consistently identifies as part of the GLM-4 series and maintains clear versioning (GLM-4.7 vs 4.6). It is transparent about its specific focus on 'agentic coding' and 'thinking' modes. There are no documented cases of the model claiming to be a competitor's product or misrepresenting its fundamental nature as an AI developed by Z.ai.

The model weights are released under the MIT license, which is a clear and permissive open-source license. However, there is some ambiguity regarding the 'governing terms' mentioned in third-party distributions (like NVIDIA NIM) and whether the training data or specific agentic frameworks used alongside the model are subject to the same permissive terms.

VRAM requirements are provided for the 'Flash' variant (16GB-24GB) and quantization (FP8) is supported. For the full 358B model, documentation suggests the use of vLLM and SGLang with multi-GPU setups (TP size 4 or 8), but detailed memory scaling for the 200K context window and specific quantization-accuracy trade-offs for the flagship version are less thoroughly documented than for the smaller variants.

Z.ai uses a clear versioning scheme (4.5, 4.6, 4.7) and maintains a basic changelog in blog posts. However, the 'silent' nature of updates to the underlying API and the lack of a detailed, granular version history for weight checkpoints make it difficult to track subtle behavior drift or performance changes over time.