---
title: "GLM 5.3 Hardware Requirements: GPU, Memory, and Deployment (2026)"
slug: glm-5-3-hardware-requirements-gpu-memory-2026
description: "GLM 5.3's open weights are 753B parameters, about 744 GB in FP8 and double that in BF16. The node configs that run it, the INT4 route, the license catch, and when to call the API instead."
author: "Yotta Labs"
date: 2026-09-21
categories: ["Inference"]
canonical: https://www.yottalabs.ai/post/glm-5-3-hardware-requirements-gpu-memory-2026
---

# GLM 5.3 Hardware Requirements: GPU, Memory, and Deployment (2026)

![](https://cdn.sanity.io/images/wy75wyma/production/44808f326c498a0ac3193d4cfdf7d67cc48ae1a5-1200x627.png)

*Z.ai released GLM 5.3’s weights two weeks after the API. Here’s what it takes to run the flagship yourself, the memory math, and the one thing that changed from GLM 5.2.*

GLM 5.3 is the same base model as GLM 5.2, so the hardware answer is mostly the one we gave for 5.2: about 753 billion parameters, roughly 40 billion active per token, around 744 GB of weights in FP8, and an 8x H200 node as the realistic serving floor. What’s new is the post-training, which is where all of 5.3’s gains come from, and the license, which is no longer MIT. Both matter if you’re deciding whether to rent a node for it.

## The numbers that matter

<!-- unsupported block: table -->

The MoE rule applies as always: 40B active parameters make the model cheap to run per token, and all 753B still have to sit in GPU memory. Low active count saves compute, not VRAM. That’s why a model that serves at Flash-class speeds needs Pro-class hardware.

## Why it’s an 8-GPU node

The FP8 checkpoint is about 744 GB. Eight H100s (640 GB) don’t hold it. Eight H200s (1,128 GB) hold it with roughly 380 GB left for KV cache, activations, and batching, which is the setup Z.ai’s own serving guidance for the GLM 5 family targets and the one our [GLM 5.2 vLLM deployment guide](https://www.yottalabs.ai/post/how-to-deploy-glm-5-2-with-vllm-on-yotta-gpu-pods) walks through on Yotta GPU Pods. Because 5.3 shares 5.2’s base and shape, that guide’s hardware sizing carries over unchanged.

<!-- unsupported block: table -->

The BF16 checkpoint at 1.5 TB is for fine-tuning and research, not serving. If you’re renting for inference, FP8 is the target, and the [pricing page](https://www.yottalabs.ai/pricing) has 8-GPU H200 and B300 nodes by the hour.

## The INT4 route

Community quantizations appeared quickly; Hugging Face lists dozens of quantized variants, including GGUF builds for llama.cpp and Ollama. At a flat 4 bits, 753B parameters is roughly 372 GB before overhead, which moves the model from an 8x H200 node down to 4x H200 (564 GB) or 8x H100 (640 GB). That’s half the hardware for a quality cost you should measure on your own tasks; for coding and agent loops it’s often acceptable, for anything where the model’s top-end reasoning is the point, less so. It never reaches a single card or a workstation. If you want GLM 5.3 quality on one GPU, the answer is the API.

## The license changed

GLM 5.2 shipped under MIT. GLM 5.3 Flash shipped under MIT. GLM 5.3 ships under a custom “glm-5.3” license on Hugging Face. We haven’t seen a plain-language summary from Z.ai of what it restricts, so before you build a hosted product on the weights, redistribute them, or fine-tune and ship the result, read the license file in the repo. For internal use and evaluation this is unlikely to matter; for anything commercial it’s the first thing to check, and it’s a real difference from every other open GLM 5 checkpoint.

## Engines and the launch command

Z.ai’s official serving paths are vLLM, SGLang, TokenSpeed, KTransformers, Transformers, and Unsloth, plus vLLM-Ascend, xLLM, and SGLang for Huawei Ascend. Docker images exist for both vLLM and SGLang.

The vLLM shape on an 8x H200 node, following the GLM 5 family recipe:

```bash
vllm serve zai-org/GLM-5.3 \
  --tensor-parallel-size 8 \
  --kv-cache-dtype fp8 \
  --max-model-len 262144 \
  --tool-call-parser glm47 \
  --reasoning-parser glm45 \
  --enable-auto-tool-choice \
  --served-model-name glm-5.3

```

--tensor-parallel-size 8 shards the weights across the node. --kv-cache-dtype fp8 matters more here than on most models because of the 1M window: the cache grows with context, and keeping it in FP8 is what leaves room for batching. --max-model-len is yours to set; Z.ai evaluates at 300K to 400K, and capping it at what your workload actually uses is the single biggest memory win. The parser flags wire up GLM’s tool calling and thinking mode; the exact parser names are pinned in the vLLM recipe for your engine version, so confirm them there rather than copying from a blog post. SGLang’s equivalent is a standard launch_server with --tp 8, and the [GLM 5.2 SGLang guide](https://www.yottalabs.ai/post/how-to-deploy-glm-5-2-with-sglang-on-yotta-gpu-pods) covers that path on the same hardware. If you’re choosing between them, the [vLLM vs SGLang comparison](https://www.yottalabs.ai/post/vllm-vs-sglang-which-inference-engine-should-you-use-in-2026) has the tradeoff.

Send reasoning_effort per request. The default is max, which is what Z.ai benchmarks with and the slowest, most expensive setting; low for routine calls and max for the hard ones is how you keep a self-hosted node’s throughput up.

## GLM 5.3 vs GLM 5.3 Flash: which one to host

Flash is 320B parameters at about 306 GiB in FP8, also an 8-GPU node but with far more headroom, and it’s MIT-licensed and multimodal. The flagship is 753B at 744 GB, text-only, custom license, and on Z.ai’s own tables it leads on the hardest agentic and knowledge benchmarks (Terminal-Bench 2.1 88.2, DeepSWE 66.9, GPQA 88.1) while Flash matches it on several agent tasks at a tenth of the API price. The [GLM 5.3 vs GLM 5.3 Flash breakdown](https://www.yottalabs.ai/post/glm-5-3-vs-glm-5-3-flash-benchmarks-which-to-use-2026) covers when the flagship earns its cost. For hosting: if Flash’s quality clears your bar, it’s the same node with more than twice the cache headroom and no license question. The [GLM 5.3 Flash hardware requirements](https://www.yottalabs.ai/post/glm-5-3-flash-hardware-requirements-gpu-memory-2026) post has that side.

## The route most teams should take

An 8x H200 node has to serve a lot of sustained traffic before it beats per-token pricing, and GLM 5.3 is on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) at $1.40 per million input tokens and $4.40 output, flat, behind the same OpenAI-compatible API as GLM 5.2, GLM 5.3 Flash, DeepSeek V4, Qwen 3.8, Kimi K3, and Grok 4.6. Self-host when data can’t leave your infrastructure, when you’re fine-tuning (and have checked the license), or when your volume is high and steady enough that owning the node beats the meter. Otherwise, call it, and use the [GLM 5.3 what’s new post](https://www.yottalabs.ai/post/glm-5-3-whats-new-benchmarks-how-to-access-2026) to decide whether the upgrade from 5.2 is worth a model-string change.

## Frequently asked questions

**How much VRAM does GLM 5.3 need?** About 744 GB for the FP8 checkpoint, which means an 8x H200 node (1,128 GB) or an 8-GPU Blackwell node. BF16 is roughly 1.5 TB and needs 8x B300 or two nodes. INT4 community builds run around 372 GB on 4x H200 or 8x H100.

**Can GLM 5.3 run on 8x H100?** Not at FP8; 640 GB is about 100 GB short of the weights. An INT4 build fits with room to spare.

**Is GLM 5.3 open source?** The weights are public on Hugging Face in BF16 and FP8, but under a custom “glm-5.3” license rather than the MIT license GLM 5.2 and GLM 5.3 Flash use. Read it before commercial use.

**Is there a GGUF or Ollama version of GLM 5.3?** Community GGUF quantizations exist, but even 4-bit builds are in the 370 GB range. It’s multi-GPU regardless of format.

**How does GLM 5.3 compare to GLM 5.2 on hardware?** Identical. Same base model, same parameter count, same checkpoint sizes, same serving recipe. If you already run 5.2, 5.3 is a checkpoint swap. The differences are post-training quality and the license.

**Which engines support GLM 5.3?** vLLM, SGLang, TokenSpeed, KTransformers, Transformers, and Unsloth on NVIDIA, plus vLLM-Ascend, xLLM, and SGLang on Ascend. Docker images are available for vLLM and SGLang.

**How does it compare to DeepSeek V4.1 Flash and Kimi K3?** GLM 5.3 at 744 GB sits between them: DeepSeek V4.1 Flash is 510 GB and needs about 614 GB of GPU memory, per the [V4.1 Flash hardware breakdown](https://www.yottalabs.ai/post/deepseek-v4-1-flash-hardware-requirements-gpu-memory-2026), and Kimi K3 is a 1.56 TB cluster-scale deployment, per the [Kimi K3 hardware post](https://www.yottalabs.ai/post/kimi-k3-hardware-requirements-gpu-memory-2026). All three are 8-GPU-or-more territory.

**What does GLM 5.3 cost on an API?** $1.40 per million input and $4.40 output on Yotta AI Gateway, flat with no peak windows. Z.ai’s own rate card may differ; check it before budgeting.

## Bottom line

GLM 5.3’s hardware story is GLM 5.2’s: 753B parameters, 744 GB in FP8, an 8x H200 node with tensor parallel 8 and an FP8 KV cache, and an INT4 route to half the hardware at a measurable quality cost. What’s new is the reason to want it (post-training gains on the hardest benchmarks) and the reason to read carefully before you deploy it (the license). If you self-host, the [8-GPU H200 and B300 nodes](https://www.yottalabs.ai/pricing) on Yotta GPU Pods are the configs above; if you’d rather not, GLM 5.3 is on [Yotta AI Gateway](https://www.yottalabs.ai/ai-gateway) one API key away.