GLM-5.2

GLM‑5.2 is Z.ai’s flagship large language model with 753 billion total parameters, of which roughly 39 billion are activated per token. The model is purpose‑built for long‑horizon, multi‑step tasks and represents a significant leap over the previous GLM‑5.1 version. For the first time in the GLM‑5 series, the model delivers stable operation with a context window of one million tokens while substantially improving computational efficiency. Architecturally, GLM‑5.2 is based on the same Mixture‑of‑Experts (MoE) structure with the DSA (DeepSeek Sparse Attention) mechanism as its predecessors, but it introduces the innovative IndexShare technology as well as refined multi‑token prediction (MTP) layers. These advances achieve a new balance among massive knowledge capacity, complex step‑by‑step reasoning, inference speed, and the quality of processing ultra‑long contexts.

The key architectural improvement in GLM‑5.2 is the IndexShare technology, which is directly responsible for the efficiency gains when handling long sequences. DSA employs a lightweight “lightning indexer” that selects only the top‑k most relevant tokens for each query, reducing the computational complexity of the core attention mechanism. However, the indexer itself requires O(L²) operations in every layer. IndexShare relies on a simple but important observation: neighboring transformer layers tend to produce similar attention patterns, meaning computing a separate index for each layer is redundant. Layers are therefore grouped: only a small number of “Full” layers perform independent indexer work, while all remaining “Shared” layers simply reuse the indices from the nearest Full layer. This cuts indexer computations by 75%, and at a context length of one million tokens it reduces FLOPs per token by a factor of 2.9, with virtually no loss in quality. In addition, the Multi‑Token Prediction (MTP) layer for speculative decoding has been improved in GLM‑5.2, increasing the length of pre‑predicted tokens by up to 20%, which further accelerates generation.

Benchmark results show that GLM‑5.2 noticeably improves upon previous versions, proves competitive with the leading closed models, and confidently leads among open solutions. On the AIME 2026 math competition, the model scores 99.2%, surpassing GPT‑5.5 (98.3%), Claude Opus 4.8 (95.7%), and Gemini 3.1 Pro (98.2%). On the CritPt critical thinking benchmark, GLM‑5.2 shares first place with Claude Opus 4.8 (20.9 points), vastly outperforming the previous GLM‑5.1 (4.6). In tool‑augmented mode on Humanity’s Last Exam (HLE w/ Tools), the model scores 54.7, beating GPT‑5.5 (52.2) and Gemini 3.1 Pro (51.4) and trailing only Claude Opus 4.8 (57.9). In coding benchmarks, GLM‑5.2 achieves 62.1% on SWE‑bench Pro, 48.9% on NL2Repo, and 81.0% on Terminal‑Bench 2.1. On the MCP‑Atlas tool‑use benchmark, the model scores 76.8, on par with Claude Opus 4.8 (77.8) and ahead of Gemini 3.1 Pro (69.2). The greatest successes of GLM‑5.2 come in benchmarks that assess the ability to handle tasks lasting hours or even days. For example, on FrontierSWE, which tests the ability to carry out work in large projects involving system optimization, large‑scale build engineering, or applied machine learning, the model scores 74.4%, trailing only Claude Opus 4.8 by less than 1%. All these figures compellingly demonstrate that the one‑million‑token context in GLM‑5.2 is not just a technical checkbox but a genuinely working tool for complex engineering tasks.

The application areas of GLM‑5.2 cover virtually every scenario that demands deep contextual understanding and long‑term planning. In software development, it can handle the full cycle – from requirements analysis to the implementation of complex projects, including writing compilers, optimizing operating‑system kernels, and creating high‑load services. In scientific research automation, it is suitable for system optimization, applied ML research, and other labor‑intensive engineering tasks. The model supports effort level control, allowing flexible tuning of the trade‑off between answer quality and response time depending on the task. Moreover, its outstanding mathematics results and scientific knowledge make GLM‑5.2 an excellent tool for education, intelligent assistants, research work, and any other domain requiring precise logical inference. Thanks to the MIT license and full openness, the model becomes an ideal choice for both commercial deployment and academic experimentation without any restrictions.

Announce Date: 16.06.2026
Parameters: 754B
Experts: 256
Activated at inference: 39B
Context: 1049K
Layers: 78
Attention Type: DeepSeek Sparse Attention
Developer: Z.ai
Transformers Version: 5.12.0
vLLM Version: >= 0.23.0
License: MIT

Use our pre-built public endpoints for free to test inference and explore GLM-5.2 capabilities. You can obtain an API access token on the token management page after registration and verification.

Rent your own physically dedicated instance with hourly or long-term monthly billing.

We recommend deploying private instances in the following scenarios:

• maximize endpoint performance,
• enable full context for long sequences,
• ensure top-tier security for data processing in an isolated, dedicated environment,
• use custom weights, such as fine-tuned models or LoRA adapters.

Contact our dedicated neural networks support team at nn@immers.cloud or send your request to the sales department at sale@immers.cloud.