Can I run GLM-5.3 on an RTX 4090?
Key facts
No: it needs about 470 GB and this card has 24 GB of graphics memory.
It needs about 469.6 GB at standard quality (estimate). Checked . Unsloth UD-Q4_K_XL GGUF.
Updated · GLM-5.3 on every card
How much memory it needs
That is about 20.1 times what this card has free.
What runs GLM-5.3 instead
Nothing among the 43 graphics cards, PCs and Macs on our list runs it fully.
See our picks for GLM-5.3 in Find hardware. Add your budget to see cheaper machines.
For comparison, an RTX 4090 costs $2,446€2,185, used (median of sold listings). Card details
GLM-5.3 on other graphics cards
Doesn't fit10
Not enough graphics memory: it needs about 469.6 GB.
- RTX 3060 12GBhas 12 GB
- RTX 4060 Ti 16GBhas 16 GB
- RTX 3090has 24 GB
- RTX 4090has 24 GB
- RTX 5080has 16 GB
- RX 9060 XT 16GBhas 16 GB
- RTX 5090has 32 GB
- RX 7900 XTXhas 24 GB
- Dual RTX 3090has 48 GB
- Dual RTX 4090has 48 GB
More about the model itself, what it is good at and its licence: GLM-5.3.
THE FIT LIST · FREE
Will the next model beat GLM-5.3 on your 128 GB? We'll tell you when it lands.
GLM-5.3 needs 469.6 GB at Q4. When a new open model would fit 128 GB, or the cards and Macs that run it change price, The Fit List tells you, with the memory it needs and dated EU/US prices. Plus a free GPU buying cheat sheet (PDF) now.
The numbers behind this answer
A model comes in several sizes. Q4 is the standard one most people use; the bigger ones answer a little better but need more memory. Each row adds up the model file, the memory a conversation takes, and some working space, and compares the total with what is free on this card.
Swipe the table sideways to see every column.
| Quality | Conversation | Result | Model file | Conversation memory | Working space | Total needed | Free memory |
|---|---|---|---|---|---|---|---|
| Q4 · standard (smallest that keeps quality) | ~12 pages8,192 tokens | Doesn't fit | 467.3 GBmeasured file | 1.5 GB | 0.8 GB | 469.6 GB | 23.4 GB |
| Q5 · a bit sharper, a bit bigger | ~12 pages8,192 tokens | Doesn't fit | 534.9 GB | 1.5 GB | 0.8 GB | 537.1 GB | 23.4 GB |
| Q8 · near full quality, almost twice Q4 | ~12 pages8,192 tokens | Doesn't fit | 806.1 GB | 1.5 GB | 0.8 GB | 808.3 GB | 23.4 GB |
| FP16 · full quality, over three times Q4 | ~12 pages8,192 tokens | Doesn't fit | 1,506.7 GB | 1.5 GB | 0.8 GB | 1,508.9 GB | 23.4 GB |
| Q4 · standard, longer conversation | ~49 pages32,768 tokens | Doesn't fit | 467.3 GBmeasured file | 5.9 GB | 0.8 GB | 474.0 GB | 23.4 GB |
speed: not estimated for MoE yet. This model only uses part of itself for each word (a mixture of experts), and our speed estimate does not handle that yet.
Sources and method
Memory and speed on this page are our estimate, not a measured test. Checked . Unsloth UD-Q4_K_XL GGUF.
Z.ai's open-weight sparse mixture of experts (GLM-5.3 License: MIT-style and permissive, plus a Model-as-a-Service clause: MaaS businesses with more than US$10 billion revenue in the past 12 months need Z.AI's security review before commercial use). Hugging Face safetensors metadata lists 753,329,940,480 parameters (F8_E4M3 751.23B, BF16 2.10B), stored here as 753.33B, so our count is an estimate. About 40B are active per token (Artificial Analysis lists 753B total / 40B active; Unsloth's guide says 744B). Config: GlmMoeDsaForCausalLM, 78 layers (first 3 dense), hidden size 6144, 64 attention heads, 256 routed experts plus 1 shared, 8 routed per token, 1 MTP layer, text only, max_position_embeddings 1,048,576. Attention is MLA (kv_lora_rank 512 + qk_rope_head_dim 64 = 576 cached values per token per layer) with a DeepSeek-style sparse indexer (32 index heads x 128, top-k 2048). We store kvHeads 1 and headDim 576, so our K+V formula gives 1.47 GB at 8,192 tokens. A runtime that caches only the 576-wide latent needs about half (0.74 GB in BF16); the indexer key cache is not counted. Our figure is a conservative estimate. Q4 uses the measured Unsloth UD-Q4_K_XL file, 467.29 GB (467,289,116,837 bytes). Q5, Q8 and FP16 are our bytes-per-parameter estimates (0.71 / 1.07 / 2.0 x 753.33B). Unsloth GGUF sizes, shards summed from the Hugging Face tree, decimal GB, checked 29 Oct 2026 (https://huggingface.co/unsloth/GLM-5.3-GGUF): UD-IQ1_S 216.72, UD-IQ1_M 228.49, UD-IQ2_M 238.58, UD-Q2_K_XL 253.88, UD-IQ3_XXS 281.69, UD-Q3_K_XL 342.97, UD-IQ4_XS 365.31, UD-Q4_K_XL 467.29, UD-Q5_K_XL 562.47, UD-Q6_K_XL 684.37, Q8_0 801.36, BF16 1,507.99. Smallest full-model file we found: antirez/glm-5.3-gguf IQ2_XXS mix, 211.08 GB (https://huggingface.co/antirez/glm-5.3-gguf). MLX: mlx-community/GLM-5.3-4bit 418.32 GB; pipenetwork/GLM-5.3-MLX-mixed-3_6bit 332.58 GB. NVIDIA NVFP4 checkpoint 464.14 GB (https://huggingface.co/nvidia/GLM-5.3-NVFP4). Official FP8 weights are 755.63 GB; BF16 repo zai-org/GLM-5.3-BF16 is 1,506.67 GB. Pruned derivatives with experts removed are not GLM-5.3 and are not scored. Runs in mainline llama.cpp per Unsloth's guide (https://unsloth.ai/docs/models/glm-5.3). Unsloth says UD-IQ2_M works on 256 GB RAM devices such as 2x DGX Spark or a Mac Studio; that is a vendor claim, not tested by us, and it relies on RAM offload. With 192 GB usable on a 256 GB Mac Studio M5 Ultra, even the 211.08 GB file does not fit, so no machine on our list runs GLM-5.3 fully in fast memory. Released by Z.ai on 18 August 2026 (API; docs.z.ai release notes and Artificial Analysis), with the Hugging Face repo created 25 August 2026. Artificial Analysis Intelligence Index v5.0: 32 for GLM-5.3 (max), checked 29 Oct 2026 (https://artificialanalysis.ai/models/glm-5-3). Thinking is always on (low / high / max). Ollama tag: not checked.
Questions
- Can I run GLM-5.3 on an RTX 4090?
- No: it needs about 470 GB and this card has 24 GB of graphics memory.
- How much memory does GLM-5.3 need?
- About 469.6 GB at the standard quality (Q4) for a conversation of up to about 12 pages of text (8,192 tokens), by our estimate: 467.3 GB for the model file (measured), 1.5 GB to hold the conversation and 0.8 GB of working space.
- How fast is GLM-5.3 on an RTX 4090?
- We don't give a speed for this model yet. This model only uses part of itself for each word (a mixture of experts), and our speed estimate does not handle that yet.
Speeds and memory are our estimate from the hardware specs, not a measured test. Prices are dated and may change. How we calculate