GLM-5.3

GLM-5.3 is the flagship model of the GLM-5 family from Z.ai. The model is built on the foundation of GLM-5.2: DSA attention, Mixture-of-Experts with 744 billion total parameters and 40 billion active per token, a context length of 1,048,576 tokens, and an MTP layer for speculative decoding to support efficient inference. The key feature of this release is that all improvements were achieved exclusively through post-training scaling. The developers significantly expanded the set of environments for long-horizon tasks, diversified training scenarios, and increased the computational resources used for training. This approach led to significant progress in solving complex programming tasks and long-running agentic scenarios, and opened new capabilities in the field of cybersecurity. Notably, unlike GLM-5.2, the main model weights are distributed in exact FP8 format, which reduces memory requirements.

GLM-5.3 uses IndexShare — an architectural innovation also inherited from GLM-5.2. This technology eliminates the bottleneck of sparse attention: the expensive indexer that computes the top-k relevant tokens at each layer with O(L²) complexity. To solve this problem, the developers observed that neighboring transformer layers form similar attention patterns, so computing them for every layer is unnecessary. IndexShare groups layers together: only a small number of “full” layers perform all indexer computations, while all other “shared” layers reuse indices from the nearest full layer. This reduces indexer computation by up to 75% even with a context of 1 million tokens.

According to benchmark results, GLM-5.3 demonstrates excellent performance, especially in programming and long-horizon agentic tasks. On Terminal Bench 3.0, the model scored 28.3%, significantly outperforming GLM-5.2 (4.6%), and on DeepSWE v1.1 it achieved 66.9% versus 46.2% for the previous version. The model ranks first (open-source SOTA) among open-weight models on the Terminal Bench 3.0 and Agents' Last Exam benchmarks. CyberGym (84.5%) deserves special attention, where GLM-5.3 surpasses all compared models, including Kimi K3, DeepSeek-V4 Pro, GPT-5.6 Sol, and Fable 5. On the ExploitBench vulnerability exploitation benchmark, the model scores 54.4%, more than double GLM-5.2 (24.4%). AutomationBench (48.2% vs. 26.2% for GLM-5.2) and HLE w/ Tools (62.5%) also confirm the model's superiority in automation and tool-use tasks.

GLM-5.3 is optimized for long-horizon tasks where the model must work autonomously for hours and iteratively refine its strategy — primarily agentic software development in the terminal and IDE (Claude Code, ZCode, OpenCode): refactoring large repositories, solving SWE tasks with real bug reports, generating entire repositories from textual descriptions (NL2Repo), where the 1M token context allows keeping the entire project in memory. A closely related use case is ML infrastructure and research tasks, where the model diagnoses bottlenecks in training stacks, runs experiments, and optimizes pipelines. Cybersecurity and vulnerability research stand out separately — white-box vulnerability discovery in source code, building exploit chains, and auditing open-source dependencies. For most code-related tasks, the developers recommend reasoning_effort=max, while for fast interactive chat scenarios — low or high with clear_thinking=true.

Announce Date: 25.08.2026
Parameters: 754B
Experts: 256
Activated at inference: 40B
Context: 1049K
Layers: 78
Attention Type: DeepSeek Sparse Attention
Developer: Z.ai
Transformers Version: 5.15.0
vLLM Version: >=0.28.0
License: GLM-5.3 License

Use our pre-built public endpoints for free to test inference and explore GLM-5.3 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 support@immers.cloud or send your request to the sales department at sale@immers.cloud.