GLM-5.3-Flash hardware requirements: VRAM, GPU and RAM
- Page updated
- Sources checked
- File sizes from Hugging Face, read October 9 and unchanged on October 17, 2026
- Memory needs as Unsloth publishes them; speeds only as published, with their setup
- Jump to the checker ↓
GLM-5.3-Flash is Z.ai's open-weight mixture-of-experts model: 320B parameters with 18B active per token, natively multimodal (text, images, video), with a 1M-token context and an MIT license. It mixes sparse and linear attention, and is the smaller sibling of GLM-5.3; its Hugging Face repository dates from August 25, 2026.
Short answer: with llama.cpp and Unsloth's GGUF files you need about 100 GB of memory in total (system RAM plus VRAM, or unified memory) for 1-bit, 115 GB for 2-bit, 128–150 GB for 3-bit and 162–210 GB for 4-bit, by Unsloth's table. No single consumer GPU holds even the smallest file (93.1 GB); on a PC most of the model sits in system RAM. Unsloth says its 3-bit UD-IQ3_XXS fits 128 GB devices such as a Mac or DGX Spark; that is its claim, not a published run, and leaves little room for context (a Mac also needs its GPU wired-memory limit raised, per the MLX LM docs). The official FP8 weights are 328.3 GB (306 GiB): KTransformers runs them with one RTX 40 or 50 card (its tutorial gives no VRAM minimum), at least 350 GB of free system memory and a CPU with AVX-512; vLLM and SGLang target server GPUs (H100 and newer, AMD Instinct MI350X/MI355X).
Which GLM-5.3-Flash GGUF does your machine reach?
Open in a full page ↗Loading private checker…
- Developer
- Z.ai Official
- Params
- 320B total, 18B active per token Official
- Context
- 1,048,576 tokens Official
- Official weights
- FP8, 328.3 GB Official
- License
- MIT Official
- llama.cpp
- with llama.cpp and Unsloth's GGUF files Community
- KTransformers
- KTransformers runs them with one RTX 40 or 50 card (its tutorial gives no VRAM minimum), at least 350 GB of free system memory and a CPU with AVX-512 Official
- vLLM / SGLang
- vLLM and SGLang target server GPUs (H100 and newer, AMD Instinct MI350X/MI355X) Official
- Ollama
- Only as glm-5.3-flash:cloud, which runs on Ollama's servers, not on your hardware Official
How much VRAM does GLM-5.3-Flash need?
File sizes as Hugging Face lists them (all parts of each file, decimal GB, read October 9, 2026; unchanged on October 17). Top-1 is how often the file picks the same next token as BF16, as measured by Unsloth.
| Format | File size | Top-1 vs BF16 | Runs with | Source |
|---|---|---|---|---|
| BF16 (official weights) | 642.7 GB | the reference | Transformers, vLLM | Z.ai |
| BF16 (GGUF) | 641.6 GB | the reference | llama.cpp | Unsloth |
| Q8_0 | 341.0 GB | not listed | llama.cpp | Unsloth |
| FP8 (official weights, the default) | 328.3 GB | not listed | vLLM, SGLang, KTransformers | Z.ai |
| UD-Q6_K_XL | 291.8 GB | 95.2% | llama.cpp | Unsloth |
| UD-Q5_K_XL | 240.3 GB | 94.4% | llama.cpp | Unsloth |
| UD-Q4_K_XL (dynamic 4-bit) | 199.7 GB | 92.2% | llama.cpp | Unsloth |
| UD-IQ4_XS | 156.8 GB | 88.2% | llama.cpp | Unsloth |
| UD-Q3_K_XL | 147.5 GB | 86.3% | llama.cpp | Unsloth |
| UD-IQ3_XXS | 120.4 GB | 81.6% | llama.cpp | Unsloth |
| UD-Q2_K_XL | 108.7 GB | 78.3% | llama.cpp | Unsloth |
| UD-IQ2_XXS | 101.8 GB | 76.3% | llama.cpp | Unsloth |
| UD-IQ1_M | 97.6 GB | 73.1% | llama.cpp | Unsloth |
| UD-IQ1_S | 93.1 GB | 70.9% | llama.cpp | Unsloth |
The vision file (mmproj) is 1.2 GB at BF16 (1.1 GB at F16). vLLM's recipe counts the FP8 checkpoint at about 306 GiB (the same 328.3 GB file) before runtime and KV-cache overhead, and the BF16 one at roughly twice that. NVFP4 builds (RedHatAI, NVIDIA, RadixArk) are for NVIDIA Blackwell GPUs with vLLM or SGLang.
What hardware can run GLM-5.3-Flash?
Unsloth's table of total memory (RAM + VRAM, or unified memory) for its GGUF files with llama.cpp. These are Unsloth's figures, not a RigCheck estimate; the guide does not say at what context length.
| Quantization | Total memory | Unsloth files of this bit width | Source |
|---|---|---|---|
| 1-bit | 100 GB | UD-IQ1_S (93.1 GB), UD-IQ1_M (97.6 GB) | Unsloth guide |
| 2-bit | 115 GB | UD-IQ2_XXS (101.8 GB), UD-Q2_K_XL (108.7 GB) | Unsloth guide |
| 3-bit | 128–150 GB | UD-IQ3_XXS (120.4 GB), UD-Q3_K_XL (147.5 GB) | Unsloth guide |
| 4-bit | 162–210 GB | UD-IQ4_XS (156.8 GB), UD-Q4_K_XL (199.7 GB) | Unsloth guide |
| 8-bit | 350 GB | Q8_0 (341.0 GB) | Unsloth guide |
| BF16 | 650 GB | BF16 (641.6 GB) | Unsloth guide |
Unsloth uses the 3-bit UD-IQ3_XXS in its demos "as it fits on 128GB devices" and says 3-bit "works on 128GB devices like a Mac or NVIDIA DGX Spark"; it publishes no run or speed on either, and the 120.4 GB file leaves little of a 128 GB machine for the OS, the context and buffers. Its introduction says the 1-bit model runs on 102 GB of RAM/VRAM, its table 100 GB. The table has no 5- or 6-bit figure (UD-Q5_K_XL 240.3 GB, UD-Q6_K_XL 291.8 GB). Image input adds the 1.2 GB vision file (mmproj); Unsloth does not say whether its table includes it. On a PC, llama.cpp splits the model between the GPU and system RAM, so a 24 or 32 GB card still needs most of the total in RAM; Unsloth Desktop does that split automatically. These are capacity figures, not tested setups for your hardware.
Which runtimes run GLM-5.3-Flash locally?
| Runtime | Status (checked October 9, 2026; Ollama, LM Studio, llama.cpp and KTransformers re-checked October 17) | Hardware it names | Source |
|---|---|---|---|
| llama.cpp | Unsloth's GGUF card and guide say to use its llama.cpp pull request / branch (which it says also adds faster decoding and MTP). llama.cpp master registers a glm5-next architecture, but we have not confirmed that master loads Unsloth's files: use Unsloth's build until it says otherwise | CUDA, CPU-only or Apple Metal builds | Unsloth guide, llama.cpp |
| Unsloth Desktop | Runs Unsloth's GGUF files, offloading to RAM automatically | macOS, Windows, Linux | Unsloth guide |
| KTransformers | Reads the official FP8 weights (about 306 GiB) directly; GPU + CPU expert inference | NVIDIA RTX 40 or 50 series (SM89 / SM120; no VRAM minimum given), an AVX-512 FP8 CPU expert kernel, at least 350 GB of available system memory | KTransformers tutorial |
| vLLM | FP8 (default), BF16 and NVFP4 checkpoints; the recipe says to use its Docker image until the integration is in the public repository, and a ROCm nightly image on AMD | NVIDIA Hopper and newer, AMD Instinct gfx950 (MI350X / MI355X); NVFP4 needs Blackwell | vLLM recipe |
| SGLang | v0.5.20 or later; recipes for FP8, MXFP4 (MI355X) and NVFP4 (Blackwell) | H100, H200, B200, B300, GB200, GB300, MI355X | SGLang cookbook |
| Ollama | Only glm-5.3-flash:cloud, which runs on Ollama's servers; no local build in the library | not applicable (cloud) | Ollama library |
| LM Studio | Its model page lists GLM-5.3-Flash only as a cloud model (LM Studio Bionic) | not applicable (cloud) | LM Studio |
Z.ai's model card also names Transformers and TokenSpeed.
How fast is GLM-5.3-Flash locally?
The only local speed we found is Unsloth's own, on a data-center GPU; we do not convert it to other hardware.
| Hardware | Setup | Generation rate | Source |
|---|---|---|---|
| 1× NVIDIA B200 | llama.cpp (Unsloth's branch), UD-IQ1_S | 32 tokens generated (tg32): 63.1 tok/s at an empty context, 59.5 at 4K, 58.0 at 16K, 49.0 at 64K. Separate MTP test, same GPU and file, 4,096-token prompt, generation length not given: 86.5 tok/s with 2 draft tokens vs 58.6 with MTP off | Unsloth guide |
We found no published speed for a consumer GPU with the model in system RAM, for a Mac or for a DGX Spark.
Measured GLM-5.3-Flash on your own hardware, or tried and it failed? Send us the numbers: paste the benchmark output or log into the form (or email it). After checking it we add it to this page, labeled as a community report and shared under CC BY 4.0.
GLM-5.3-Flash specs
| Parameters | 320B total, 18B active per token Model card |
|---|---|
| Architecture | 45 layers: 34 linear-attention (KDA) and 11 sparse-attention (DSA) layers; mixture of experts with 288 routed experts (8 per token) plus 1 shared expert; one MTP layer for speculative decoding config.json, KTransformers tutorial |
| Context length | 1,048,576 tokens config.json |
| Input | Text, images and video config.json, vLLM recipe, KTransformers tutorial |
| Official weights | FP8, 328.3 GB (the default repository); BF16, 642.7 GB Hugging Face |
| License | MIT Model card |
Common setups
The checker's own answer for 8 common setups (graphics cards with 32 or 64 GB of system RAM, then Macs), in the finder's quick mode (published recipes and memory tables at their own context; Unsloth's GLM table states no context length). Answers are published route matches, not measured setups; "Unknown" means the checker has no route or source for that hardware. Your own setup: the checker.
| Setup | Answer |
|---|---|
| GeForce RTX 3060 (12 GB) + 32 GB RAM | ✕ Below Unsloth's smallest memory figure (100 GB for 1-bit) Unsloth's table puts the 1-bit GGUF (93.1–97.6 GB) at 100 GB of total memory, RAM + VRAM or unified memory. This setup counts 12 GB of VRAM + 32 GB of RAM = 44 GB. |
| GeForce RTX 4060 Ti (16 GB) + 32 GB RAM | ✕ Below Unsloth's smallest memory figure (100 GB for 1-bit) Unsloth's table puts the 1-bit GGUF (93.1–97.6 GB) at 100 GB of total memory, RAM + VRAM or unified memory. This setup counts 16 GB of VRAM + 32 GB of RAM = 48 GB. |
| GeForce RTX 3090 / 3090 Ti (24 GB) + 64 GB RAM | ✕ Below Unsloth's smallest memory figure (100 GB for 1-bit) Unsloth's table puts the 1-bit GGUF (93.1–97.6 GB) at 100 GB of total memory, RAM + VRAM or unified memory. This setup counts 24 GB of VRAM + 64 GB of RAM = 88 GB. |
| GeForce RTX 5090 (32 GB) + 64 GB RAM | ✕ Below Unsloth's smallest memory figure (100 GB for 1-bit) Unsloth's table puts the 1-bit GGUF (93.1–97.6 GB) at 100 GB of total memory, RAM + VRAM or unified memory. This setup counts 32 GB of VRAM + 64 GB of RAM = 96 GB. |
| Apple Silicon Mac (64 GB unified) | – Unknown No published GLM-5.3-Flash route checked here for this device. |
| Mac with M5 Max, 40-core GPU (128 GB unified) | ! Matches Unsloth's 3-bit figure (128 GB): UD-IQ3_XXS (120.4 GB) (after raising the Mac's GPU wired-memory limit) (published route match) Unsloth's total-memory table (RAM + VRAM, or unified memory) for its GGUF files with llama.cpp; it does not say at what context length, and which file of a bit width fits at 128 GB is our reading. This setup counts 128 GB of unified memory. Unsloth says 3-bit works on 128 GB devices like a Mac or DGX Spark; the 120.4 GB file leaves little room for the OS and context, and no run or speed is published. Image input adds the 1.2 GB vision file. Use Unsloth's llama.cpp branch (glm5next/upstream) or Unsloth Desktop, as Unsloth's guide says. No published speed for this hardware. On a Mac, raise the GPU wired-memory limit first. |
| Mac Studio M3 Ultra (256 GB unified) | ! Matches Unsloth's 4-bit figure (210 GB): UD-Q4_K_XL (199.7 GB) (after raising the Mac's GPU wired-memory limit) (published route match) Unsloth's total-memory table (RAM + VRAM, or unified memory) for its GGUF files with llama.cpp; it does not say at what context length, and which file of a bit width fits at 210 GB is our reading. This setup counts 256 GB of unified memory. Image input adds the 1.2 GB vision file. Use Unsloth's llama.cpp branch (glm5next/upstream) or Unsloth Desktop, as Unsloth's guide says. No published speed for this hardware. On a Mac, raise the GPU wired-memory limit first. |
| Mac Studio M3 Ultra (512 GB unified) | ! Matches Unsloth's 8-bit figure (350 GB): Q8_0 (341.0 GB) (after raising the Mac's GPU wired-memory limit) (published route match) Unsloth's total-memory table (RAM + VRAM, or unified memory) for its GGUF files with llama.cpp; it does not say at what context length, and which file of a bit width fits at 350 GB is our reading. This setup counts 512 GB of unified memory. Image input adds the 1.2 GB vision file. Use Unsloth's llama.cpp branch (glm5next/upstream) or Unsloth Desktop, as Unsloth's guide says. No published speed for this hardware. On a Mac, raise the GPU wired-memory limit first. |
Machine-readable: the same answers as JSON (CC BY 4.0, attribution "RigCheck").
Related hardware
Graphics card and Mac pages where this page's checker gives an answer, smallest memory first: graphics cards with 64 GB of system RAM, Macs with their unified memory, in the finder's quick mode (published route match; each page lists every model).
- Mac with M5 Max (128 GB): Matches Unsloth's 3-bit figure (128 GB): UD-IQ3_XXS (120.4 GB) (after raising the Mac's GPU wired-memory limit)
- Mac Studio M3 Ultra (256 GB): Matches Unsloth's 4-bit figure (210 GB): UD-Q4_K_XL (199.7 GB) (after raising the Mac's GPU wired-memory limit)
- Mac Studio M3 Ultra (512 GB): Matches Unsloth's 8-bit figure (350 GB): Q8_0 (341.0 GB) (after raising the Mac's GPU wired-memory limit)
All graphics cards · All Macs · Any other machine
FAQ
How much RAM or VRAM does GLM-5.3-Flash need?
By Unsloth's table, about 100 GB of total memory (system RAM plus VRAM, or unified memory) for its 1-bit GGUF, 115 GB for 2-bit, 128–150 GB for 3-bit, 162–210 GB for 4-bit, 350 GB for 8-bit and 650 GB for BF16. The files are 93.1 GB (UD-IQ1_S) to 641.6 GB (BF16); the official FP8 weights are 328.3 GB.
Can GLM-5.3-Flash run on an RTX 4090 or RTX 5090?
Not on the card alone: even the 93.1 GB 1-bit file is far larger than 24 or 32 GB. With llama.cpp the rest goes into system RAM, so by Unsloth's table you need about 100 GB of RAM and VRAM together for 1-bit. KTransformers runs the official FP8 weights on one RTX 40 or 50 card with the experts on the CPU, but it asks for at least 350 GB of available system memory, its CPU expert kernel is written for AVX-512, and its tutorial gives no VRAM minimum.
Can I run GLM-5.3-Flash on a Mac?
Unsloth's guide gives a Metal build of its llama.cpp branch and says its 3-bit UD-IQ3_XXS (120.4 GB) works on 128 GB devices such as a Mac, which leaves little room for context; raise the Mac's GPU wired-memory limit first (sudo sysctl iogpu.wired_limit_mb=…, above the model's size and below the Mac's memory; MLX LM docs). Its table puts 4-bit at 162–210 GB, so by our reading of that table UD-Q4_K_XL needs a 256 GB Mac. We found no published run or speed on a Mac.
Will GLM-5.3-Flash run on a DGX Spark?
Unsloth says its 3-bit UD-IQ3_XXS works on 128 GB devices like the NVIDIA DGX Spark, with little room left for context. We found no published run or speed on it.
Is GLM-5.3-Flash on Ollama?
Only as glm-5.3-flash:cloud, which runs on Ollama's servers, not on your hardware (Ollama library, checked October 17, 2026). To run it locally, use llama.cpp or Unsloth Desktop with Unsloth's GGUF files.
Does llama.cpp support GLM-5.3-Flash?
Through Unsloth's build: its GGUF card and guide say to use its llama.cpp pull request / branch, which it says also adds faster decoding at long context and MTP. llama.cpp master registers a glm5-next architecture, but we have not confirmed that master loads Unsloth's files, and Unsloth's guide still builds its own branch (checked October 17, 2026).
Which GLM-5.3-Flash GGUF should I download?
The largest Unsloth file whose total-memory row (RAM plus VRAM, or unified memory) your machine meets. Our reading of Unsloth's table, which gives ranges per bit width, not per file: UD-IQ1_S or UD-IQ1_M with about 100 GB, UD-IQ2_XXS or UD-Q2_K_XL with 115 GB, UD-IQ3_XXS (128 GB) or UD-Q3_K_XL (150 GB), UD-IQ4_XS (162 GB) or UD-Q4_K_XL (210 GB), Q8_0 with 350 GB. Unsloth's own demos use UD-IQ3_XXS.
Which GPUs does GLM-5.3-Flash need with vLLM or SGLang?
Server GPUs. vLLM supports NVIDIA Hopper and newer and AMD Instinct gfx950 (MI350X / MI355X) with the FP8 weights (about 306 GiB, the 328.3 GB file); its launch example is TP4 on one GB200 tray, its NVFP4 build needs Blackwell, and until the integration is in the public vLLM repository the recipe says to use Docker. SGLang's cookbook has recipes for H100, H200, B200, B300, GB200, GB300 and four MI355X (FP8 or MXFP4 weights).
Is GLM-5.3-Flash open source?
Yes: the weights are on Hugging Face under the MIT license.
Sources
- Hugging Face: zai-org/GLM-5.3-Flash (model card, config.json, file sizes, repository date from the Hugging Face API) and zai-org/GLM-5.3-Flash-BF16 (read October 9, 2026)
- Hugging Face: unsloth/GLM-5.3-Flash-GGUF and Unsloth's guide (GGUF sizes, top-1 accuracy, total-memory table, build options, B200 speeds)
- KTransformers tutorial, vLLM recipe and SGLang cookbook
- Ollama library: glm-5.3-flash tags, LM Studio: GLM-5.3-Flash; llama.cpp llama-arch.cpp (checked October 9, 2026); MLX LM README (raising the macOS GPU memory limit)