GLM-5.3 Weights Released: 744B MoE Giant Model with vLLM Day-0 Support

GLM-5.3 Weights Go Open Source: Zhipu Releases 744B Giant Model

Zhipu (Zai_org) has officially open-sourced the GLM-5.3 model weights, representing the largest model in the GLM-5.3 series. What makes this release particularly notable is that the popular inference framework vLLM offers Day-0 support, meaning developers can seamlessly deploy and run the model on the very day it launched.

The publicly disclosed specifications are impressive:

• Total parameters: 744B (744 billion)
• Activated parameters: 40B (40 billion)
• Context length: Up to 1M (1 million tokens)
• Maximum output: 128K tokens

The combination of 744B total parameters with only 40B activated parameters makes it clear this is a sparse MoE (Mixture of Experts) architecture. This design gives the model a massive knowledge capacity while keeping inference costs relatively manageable — only about 5.4% of parameters are activated per forward pass.

Keeping the Base, Scaling Post-Training

One notable technical detail in this release: Zhipu did not retrain the model's base. Instead, they retained the GLM-5.2 base and focused on extending and reinforcing the post-training stage.

As stated in the original announcement: "@Zai_org kept the GLM-5.2 base and scaled post-training instead, so nothing under the model changed."

There's a clear engineering rationale behind this strategy.

Advantages of Reusing the Base

Pre-training a model at the 744B scale is extraordinarily expensive, requiring massive compute and data resources. If the GLM-5.2 base already has strong foundational capabilities, investing additional effort in post-training stages — such as SFT, RLHF, and DPO — can deliver significant capability gains at a fraction of the cost.

This reflects a broader trend in the large model race: the focus is gradually shifting from "scaling pre-training" to "optimizing post-training quality." As base model capabilities approach saturation, improvements in alignment, reasoning, and instruction following during post-training become the key differentiators.

Direct Benefits for Inference and Deployment

Because the underlying architecture is unchanged, GLM-5.3 can fully reuse GLM-5.2's inference path — clearing the way for immediate toolchain support.

vLLM Day-0 Support: GLM-5.3 Inference Deployment in Detail

The vLLM team has confirmed that GLM-5.3 runs directly through GLM-5.2's inference path in vLLM, with no changes required. Specifically:

• Same parsers: Reuses the glm47 and glm45 parsers
• Same MTP: Multi-Token Prediction configuration is read directly from the checkpoint
• Same FP8 KV cache: Supports full 1M context

For developers, the deployment command is refreshingly simple:

vllm serve zai-org/GLM-5.3 -tp 8

The -tp 8 flag specifies 8-way Tensor Parallelism, which also hints at the multi-GPU hardware requirements for running a 744B model.

How FP8 KV Cache Enables Million-Token Context

FP8 KV cache support up to the full 1M context is one of GLM-5.3's key capabilities. One of the biggest bottlenecks in long-context inference is that KV cache VRAM usage grows linearly with sequence length. Storing the KV cache in FP8 precision dramatically reduces memory overhead — roughly half compared to FP16 — which is the core technical enabler for making million-token context practical in real deployments.

Combined with the MTP (Multi-Token Prediction) mechanism, GLM-5.3 is well-positioned for tasks like long-document summarization and ultra-long document analysis.

What GLM-5.3's Open Source Release Means for the Community

The release of GLM-5.3 weights, paired with Day-0 support from a mainstream inference framework, reflects a growing maturity in the open-source ecosystem around Chinese-developed large models. For developers and enterprises, this means:

1. Ready to use immediately: No waiting for community adaptation — deployable in production on launch day
2. Manageable inference costs: MoE sparse activation + FP8 optimization dramatically lowers the barrier to running an ultra-large model
3. Top-tier scale: 744B total parameters places it among the largest open-source models currently available

For enterprises and research institutions looking to run very large models on their own infrastructure, GLM-5.3 offers a compelling combination of parameter scale, long-context capability, and engineering accessibility.

Summary

The release of GLM-5.3 is less a debut of an entirely new model and more a textbook example of capability uplift driven by post-training. By reusing the GLM-5.2 base and concentrating investment in post-training, Zhipu managed to control R&D costs while ensuring seamless toolchain compatibility. vLLM's Day-0 support and FP8 long-context optimization are what make this 744B MoE model genuinely viable for real-world engineering deployment.

Note: The information above is primarily sourced from vLLM's official announcements on social platforms. Some technical details — such as specific benchmark results — are still pending official disclosure and community verification.