GLM-5.3: Benchmarks, Weights Date & GLM-5.2 Diff (2026)

On August 14, 2026, Z.ai (Zhipu AI) released a new model called GLM-5.3 via the GLM Coding Plan. It has 743B parameters, and its base model is identical to GLM-5.2 — meaning it was not pretrained from scratch, but instead had its coding and agentic capabilities extended purely through post-training. Z.ai's internal evaluations claim roughly a 50% improvement in coding ability over GLM-5.2, and the model also posts top scores among open models on public benchmarks such as Terminal-Bench 3.0 and Agent's Last Exam. However, as of release the weights are not yet available; access is currently limited to API and service channels. As of this writing, Z.ai has stated it plans to release the weights roughly two weeks after launch (around late August 2026), pending a security review.

The defining feature of GLM-5.3 is that its 743B-parameter base model itself is unchanged from GLM-5.2. Rather than modifying the pretraining corpus or model architecture, Z.ai improved coding and agentic capability solely through the post-training phase — the additional tuning stage after pretraining, which includes reinforcement learning and agent-oriented fine-tuning. We previously covered GLM-5.2's hardware requirements in GLM-5.2 Requirements and VRAM Guide, and since the base model is unchanged, the general scale of compute needed for inference should carry over. According to Z.ai's own evaluation, this post-training-only approach delivered roughly a 50% improvement in coding capability over GLM-5.2. Claiming a gain of this magnitude without retraining the base model is a notable example that pretraining scale-up is not the only lever for improving a model.

GLM-5.3 has published measured scores across several public benchmarks. Notably, it ranks first among open models on Terminal-Bench 3.0, Agent's Last Exam, and AutomationBench — all of which evaluate agentic coding and security-oriented tasks.

As the benchmark names CyberGym, ExploitBench, and ExploitGym suggest, GLM-5.3's scores skew heavily toward security and vulnerability-discovery tasks rather than general-purpose coding alone. This is directly tied to the post-training design philosophy discussed below.

In recent LLM development, scaling up pretraining — more parameters, more training tokens — has been the dominant path to better performance. GLM-5.3 demonstrates a different route. Without re-running pretraining on the base model, Z.ai claims it achieved a substantial gain in a specific domain (agentic coding and security tasks in this case) purely through additional post-training work such as reinforcement learning and task-specific fine-tuning. The advantage of this approach is clear: post-training requires far less compute and development time than pretraining. The short cycle between GLM-5.2 and GLM-5.3 likely reflects this asymmetry. At the same time, this approach implies a constraint: the base model's underlying knowledge and reasoning 'strength' does not itself change. Post-training draws out existing capability and adapts it to specific tasks; it does not create capabilities the base model never had. That GLM-5.3's gains concentrate in agentic, coding, and security-related benchmarks is consistent with this characteristic.

GLM-5.3 was post-trained using data and execution environments built specifically for vulnerability discovery. Z.ai describes this training as having led the model to 'begin reasoning about multi-stage attack chains as a coherent plan' — that is, rather than handling individual exploits in isolation, it can assemble multi-step attack scenarios into a single plan. As a demonstration of this capability, Z.ai says its security research effort discovered 2,436 vulnerabilities across 269 projects, publishing the results to the public registry cvd.z.ai. This is valuable for defenders, but the same capability carries an inherent dual-use risk, since it could equally be applied offensively. The high scores on CyberGym, ExploitBench, and ExploitGym benchmarks underscore how strong this capability is. It is reasonable to infer that the need to evaluate this offensive-capability risk is part of why weights were not released at launch, with Z.ai citing roughly two weeks for a security review before release. That said, Z.ai has not explicitly stated this as the reason, and this remains a general inference rather than a confirmed fact.

GLM-5.3 is currently accessible via the GLM Coding Plan, with integrations available for Z.ai's own coding tool ZCode as well as third-party agentic coding tools including Claude Code and OpenCode. API pricing, context length, and licensing terms could not be confirmed from official sources as of this writing, and all remain unpublished. Rather than offering speculative figures, anyone considering adoption should check Z.ai's official announcements directly.

As the table shows, GLM-5.3 is distinguished from GLM-5.2 not by model scale or architecture, but by how the training was finished, the resulting benchmark performance, and whether weights are publicly available.

As of 2026, the field of open coding and agentic models is expanding rapidly. The DeepSeek camp has emphasized pricing and harness integration with releases like DeepSeek V4 Pro 0813 GA, while the Qwen series (including Qwen 3.8) has leaned on a wide range of parameter sizes and open weight distribution. Models such as MiniMax H3 and Muse Glimmer each carve out their own niche around inference efficiency or multimodal support. Within this landscape, GLM-5.3 is not, strictly speaking, an 'open-weight' model yet — its weights have not been released — but rather a model that has announced it will become open-weight. At the same time, its security- and vulnerability-discovery focus is a differentiator that other major models have not emphasized as heavily, which could give it particular relevance for offensive security research or defensive code auditing use cases rather than general-purpose coding alone. The conditions are not uniform enough to directly rank benchmark scores across all these models, so it is more useful to think of them as differing in character depending on the intended use case.

- If you want to try it for coding/agentic use right away: API access via the GLM Coding Plan and integration with ZCode, Claude Code, and OpenCode are already available. Pricing and context length are unpublished, so check official sources first
- If you don't want to expose your own data to the model, or need on-premises/air-gapped deployment: this isn't an option yet since weights are unreleased; you'll need to wait for the planned release around late August 2026
- If your interest is security research or vulnerability investigation: the 2,436 vulnerability findings Z.ai published are viewable on cvd.z.ai without needing to run the model itself
- If you want to prepare for local deployment in parallel: our Local LLM Inference Engine Comparison covers general considerations around quantization and multi-GPU setups for models in this parameter class, which may help with planning ahead of the weight release