GLM-5.3-Flash: VRAM requirements by quant
Key facts
GLM-5.3-Flash needs about 200.7 GB at standard quality (Q4, estimate).
Checked . Unsloth UD-Q4_K_XL GGUF.
GLM-5.3-Flash is a very large open AI model. It needs about 201 GB of memory at standard quality, more than any PC or Mac we track. Models this size usually run on rented servers.
LIVE IN BROWSER · $0 · NO LOGIN
Test comparable open models (Recursive Flint 40B vs Gemma 4 31B) free in your browser
Open full 8-model playground ↗Click a starter prompt below or type your own question to watch Recursive Flint stream side by side against Gemma 4 31B right on this page:
Memory needed
About 201 GB
At standard quality, with room for a conversation of about 12 pages.
Cheapest machine that runs it
None of ours
It needs more than any PC or Mac we track.
THE FIT LIST · FREE
Will the next model beat GLM-5.3-Flash on your 128 GB? We'll tell you when it lands.
GLM-5.3-Flash needs 200.7 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.
Can a PC or Mac run GLM-5.3-Flash?
None of the PCs or Macs we track can hold GLM-5.3-Flash: it needs about 201 GB. Models this size usually run on rented servers. What renting costs
Speeds and memory are our estimate from the hardware specs, not a measured test. Prices are dated and may change. How we calculate
Memory at each quality level
A lower quality level makes the model smaller, so it fits on a smaller machine. Standard (Q4) is what most people run.
| Quality | Memory needed |
|---|---|
| Q4 · standard (smallest that keeps quality) | 200.7 GB |
| Q5 · a bit sharper, a bit bigger | 229.1 GB |
| Q8 · near full quality, almost twice Q4 | 344.8 GB |
| FP16 · full quality, over three times Q4 | 643.6 GB |
With a conversation of 8,192 tokens · about 12 pages of text.
Sources and method
At standard quality: 199.7 GB for the model itself (a measured file), 0.2 GB for a conversation of 8,192 tokens, and 0.8 GB for the software.
Size source: huggingface.co/zai-org/GLM-5.3-Flash. 321.32B parameters in total, 18B used for each word. The memory counts the total.
Z.ai's open-weight multimodal mixture of experts (MIT licence), the smaller sibling of GLM-5.3. The model card says 320B total parameters with 18B active per token. Hugging Face safetensors metadata lists 321,323,031,390 parameters (F8_E4M3 314.40B, BF16 6.93B, F32 295,518), including the vision encoder and 1 MTP layer, stored here as 321.32B, so our count is an estimate. Config (Glm5NextForConditionalGeneration, text_config glm5_next_text): 45 layers (first 3 dense), hidden size 4096, 288 routed experts plus 1 shared, 8 routed per token, 1 MTP layer, max_position_embeddings 1,048,576. Attention is hybrid: 34 linear-attention layers (KDA delta-rule linear attention, 64 heads x 128) and 11 DeepSeek-style sparse-attention layers (MLA, kv_lora_rank 512, qk_rope_head_dim 0, sparse indexer 32 heads x 128, top-k 2048), every 4th layer from layer 3. Our estimate counts conversation memory only for the 11 sparse-attention layers (latent width 512, K+V formula), 0.18 GB at 8,192 tokens. Linear-attention layers are stored with a sliding window of 0 so our estimate gives them no conversation memory; their fixed recurrent state and the indexer key cache are not counted. Q4 uses the measured Unsloth UD-Q4_K_XL file, 199.71 GB (199,707,321,472 bytes). Q5, Q8 and FP16 are our bytes-per-parameter estimates (0.71 / 1.07 / 2.0 x 321.32B). Unsloth GGUF sizes, shards summed from the Hugging Face tree, decimal GB, checked 29 Oct 2026 (https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF): UD-IQ1_S 93.09, UD-IQ1_M 97.58, UD-IQ2_XXS 101.84, UD-Q2_K_XL 108.72, UD-IQ3_XXS 120.37, UD-Q3_K_XL 147.54, UD-IQ4_XS 156.82, UD-Q4_K_XL 199.71, UD-Q5_K_XL 240.31, UD-Q6_K_XL 291.83, Q8_0 340.98, BF16 641.64; vision mmproj-F16 1.13 GB extra. ggml-org/GLM-5.3-Flash-GGUF: Q4_K 181.37. antirez/glm-5.3-flash-gguf: Q2 96.51, Q4_K 190.88. MLX: mlx-community/GLM-5.3-Flash-4bit 203.99 GB; pipenetwork/GLM-5.3-Flash-MLX-4bit 177.55 GB. NVIDIA NVFP4 checkpoint 204.44 GB. Official FP8 weights are 328.34 GB. Pruned derivatives are not GLM-5.3-Flash and are not scored. Mainline llama.cpp added GLM-5.3-Flash support in pull request 27773, merged 30 September 2026 (https://github.com/ggml-org/llama.cpp/pull/27773). Unsloth says the 1-bit quant works on 100 GB RAM and 3-bit on 128 GB devices (https://unsloth.ai/docs/models/glm-5.3-flash); that is a vendor claim, not tested by us. On our rule a 256 GB Mac Studio M5 Ultra (192 GB usable) holds UD-IQ4_XS (157.80 GB with context) but not UD-Q4_K_XL (200.69 GB); a 128 GB box (96 GB usable) holds only UD-IQ1_S (94.07 GB). Released 26 August 2026 (Artificial Analysis); Hugging Face repo created 25 August 2026. Artificial Analysis Intelligence Index v5.0: not scored yet (checked 29 Oct 2026). Thinking modes low / high / max. Ollama lists only a cloud tag (glm-5.3-flash:cloud), checked 29 Oct 2026.
Licence
Commercial use: yes
Commercial use allowed. Keep the licence and copyright notice when you share the weights.
Licence: MIT · checked 29 Oct 2026
Our reading of the licence, not legal advice.
Will GLM-5.3-Flash run on my graphics card?
Common questions
- How much memory does GLM-5.3-Flash need?
- About 200.7 GB at standard quality (Q4) with a conversation of 8,192 tokens: 199.7 GB for the model itself, 0.2 GB for the conversation and 0.8 GB for the software. The model size is a measured file; the conversation and software parts are our estimate.
- Does GLM-5.3-Flash fit on a single 24 GB graphics card?
- Not in full. It needs about 200.7 GB, and a 24 GB card leaves about 23.4 GB usable.
- Are the speeds measured?
- We do not show a speed for this model. It is a mixture-of-experts model, and our estimate would not be reliable for that design.
- Can I run GLM-5.3-Flash locally?
- Yes, on a large unified-memory machine, at a lower quant. GLM-5.3-Flash has about 320B parameters (18B active). With 8,192 tokens of context, Unsloth's UD-IQ4_XS file needs about 157.8 GB, which fits the 192 GB usable on a 256 GB Mac Studio M5 Ultra. No graphics card or dual-GPU build on our list holds it.
- How much memory does GLM-5.3-Flash need at 4-bit?
- Unsloth's UD-Q4_K_XL GGUF is 199.71 GB, so our estimate is about 200.7 GB with 8,192 tokens of context and working space. That is just over the 192 GB usable on a 256 GB Mac. Smaller 4-bit files exist: UD-IQ4_XS is 156.82 GB and ggml-org's Q4_K is 181.37 GB.
- Can GLM-5.3-Flash run on a 128 GB Mac or Strix Halo mini PC?
- Only at 1-bit. Unsloth's UD-IQ1_S file is 93.09 GB, about 94.1 GB with 8,192 tokens of context, inside the 96 GB we count as usable on 128 GB Macs and Strix Halo boxes, with under 2 GB to spare. The next file up (UD-IQ1_M, 97.58 GB) does not fit. Unsloth says 3-bit works on 128 GB devices; that relies on more memory than our rule allows and we have not tested it.
- How much context memory does GLM-5.3-Flash use?
- Very little. Only 11 of its 45 layers use attention that grows with context (sparse attention with a compressed 512-wide cache); the other 34 are linear-attention layers with a fixed-size state. Our estimate is 0.18 GB at 8,192 tokens, excluding that fixed state and the indexer cache.
- Does llama.cpp or MLX support GLM-5.3-Flash?
- Yes. Mainline llama.cpp added GLM-5.3-Flash support in pull request 27773, merged on 30 September 2026, and ggml-org, Unsloth and others publish GGUF files. MLX builds exist too, for example mlx-community/GLM-5.3-Flash-4bit at 203.99 GB and pipenetwork's 4-bit build at 177.55 GB.
- Is GLM-5.3-Flash free for commercial use?
- Yes. Z.ai publishes the weights under the MIT licence, which allows commercial use. Read the licence on the model page before you ship.