GLM 4.5 — Hardware Requirements & GPU Compatibility
ChatFunctions
GLM-4.5 is Z.ai's flagship model in the GLM-4.5 series, a foundation model built for intelligent-agent applications with 355 billion total parameters and about 32 billion active via its Mixture-of-Experts design. It is a hybrid reasoning model, offering a thinking mode for complex reasoning and tool use alongside a non-thinking mode for fast responses, and it unifies reasoning, coding, and agentic capabilities in one checkpoint alongside the smaller GLM-4.5-Air. It ships with a 128K token context window under the MIT license, permitting commercial and research use. At 355 billion total parameters, even 4-bit quantization needs roughly 200 GB of memory, which puts local use in multi-GPU server territory; most people will access it through a hosted endpoint instead.
GLM 4.5 runs on a 256 GB+ Mac or a datacenter GPU
Best local pick → Mac Studio (M3 Ultra, 256GB) · Q4_K_M · 217.9 GB · ~18 tok/s
Macs from 128 GB (IQ2_M · ~25 tok/s)
Score 100.9#54 of 106 open#38 of 89 that fit ≤256 GBMIT
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
llmrun Composite Score →
100.9
- 90% range 96.5–105.2
- #54 of 106 open models
- #38 of 89 that fit ≤256 GB
Run GLM 4.5
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/unsloth/GLM-4.5-GGUF:Q4_K_MQ4_K_M runs on one Instinct MI325X 256 GB · 16K context
Download: Q4_K_M/GLM-4.5-Q4_K_M-00001-of-00005.gguf216.46 GB + 4 partsunsloth
Ways to runOther ways
each opens step-by-stepRun in cloud
Rent 4× H100 SXM · Q4_K_M
- Estimated hourly price
- ~$9.09/hr
- est., whole server
- GPUs
- 4 GPUs
- 320 GB of VRAM
How Much VRAM Does GLM 4.5 Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | Runs on | |
|---|---|---|---|---|
| 152.9 GB | 157.7 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Runs on B200 192 GB | |
| 157.4 GB | 162.2 GB | 3-bit small — 3-bit small quantization | Runs on B200 192 GB | |
| 175.3 GB | 180.1 GB | 3-bit medium — 3-bit medium quantization | Runs on B200 192 GB | |
| 179.8 GB | 184.6 GB | 4-bit · legacy — 4-bit legacy quantization | Runs on B200 192 GB | |
| 215.6 GB | 220.4 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Runs on Instinct MI325X 256 GB | |
| 255.9 GB | 260.8 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Runs on Instinct MI350X 288 GB | |
| 296.3 GB | 301.1 GB | 6-bit, very good — 6-bit quantization, very good quality | Runs on 5× H100 SXM 80 GB | |
| 359.0 GB | 363.8 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Runs on 6× H100 SXM 80 GB |
Selected Q4_K_M
4-bit medium quantization217.9GBat 16K context
- Weights
- 215.0 GB
- KV cache
- 2.6 GB
- Overhead
- 0.3 GB
Server-class: ~217.9 GB at 16K context · one Instinct MI325X 256 GB
Which Hardware Can Run GLM 4.5?
Q4_K_M · 215.6 GBGLM 4.5 (Q4_K_M) requires 215.6 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 281+ GB is recommended. Using the full 131K context window can add up to 20.3 GB, bringing total usage to 235.9 GB. No consumer GPU has enough memory. 2 devices with unified memory can run GLM 4.5, including Mac Studio (M3 Ultra, 512GB), Mac Studio (M3 Ultra, 256GB).
128–144 GB
Q4_K_M · best fit IQ2_M (123.8 GB)
4 GPUs · 2 Mac classes
GH200 Grace Hopper Superchip~113Simulate output on GH200 Grace Hopper SuperchipH200 NVL~112Simulate output on H200 NVLH200 SXM~112Simulate output on H200 SXMM2 Ultra*128 GB~25Simulate output on Mac Pro M2 Ultra (128 GB)M5 Max128 GB~22Simulate output on Mac Studio M5 Max (128 GB)192+ GB
Q4_K_M · best fit Q3_K_L (186.5 GB)
6 GPUs · 1 Mac class · 2 device families
B200~117Simulate output on B200B300~117Simulate output on B300Instinct MI350X~45Simulate output on Instinct MI350XM3 Ultra*256–512 GB~20Simulate output on Mac Studio (M3 Ultra, 256GB)8x H100 SXM5~164Simulate output on NVIDIA DGX H1008x A100 SXM4~147Simulate output on NVIDIA DGX A100 640GB
By VRAM tier · typical GPUs and Macs at Q4_K_M
Consumer GPUs (up to 32 GB) and Macs up to 96 GB: too small for this model (needs ≥ 101.4 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →
How good is GLM 4.5?
The llmrun Score merges 7 public benchmark boards into a single figure, where 100 represents the average for open models.
Score vs VRAM in the ≤256 GB tier
- GLM 4.5
- Other models
- Best Score for its VRAM
#38 of 89 that fit ≤256 GB
Each dot is a model at the best quant that fits ≤256 GB. The line steps up to the best Score available at each amount of VRAM; a dot below it is beaten by a model that needs no more memory.
Full ≤256 GB ranking →Benchmarks
The 6 boards that weigh most in the llmrun Score, of 8 with a result for this model; badges show who ran the test.
- SWE-bench Verified
SWE-bench · Jul 2025Independent
Independent64.2%
#9/20 open
- LMArena Text
LMArena · Oct 2026Arena
Arena1410.9
#44/188 open
- Humanity's Last Exam
Epoch AI · Aug 2025Lab-reported
Lab-reported8.3%
#2/4 open
- MMLU-Pro
TIGER-LabIndependent
Independent84.6%
#10/146 open
- SWE-bench Bash Only
SWE-bench · Aug 2025Independent
Independent54.2%
#6/9 open
- WeirdML
Epoch AI · Aug 2025Independent
Independent40.6%
#23/47 open
In the ≤256 GB class
89 models fit ≤256 GBSame Score, only models that fit ≤256 GB at their best quant: the ranking a ≤256 GB owner actually chooses from.
- #1
DeepSeek V4.1 Flash 1017DeepSeek
134.4
~107 tok/s · 229.5 GB
- #2
MiMo V2.6 Flash RLXiaomi
131.9
~83 tok/s · 186.5 GB
- #3
DeepSeek V4.1 FlashDeepSeek
129.4
~108 tok/s · 210.4 GB
- #4
Step 5 PreviewStepFun
128.0
~48 tok/s · 191.4 GB
- #5
DeepSeek V4 Flash 0731DeepSeek
127.6
~76 tok/s · 183.0 GB
- #6
GLM 5.3 FlashZ.ai
124.1
~83 tok/s · 205.2 GB
Alternatives to GLM 4.5
GLM 4 family · 18 models →Official models with an llmrun Score, each answering one question; sizes are each model's default quant (usually Q4_K_M) with a 16K-token context.
Best alternative at this size
MiMo V2.6 Flash RL
Xiaomi
- 31 higher Score
- 31.4 GB less VRAM
Score 131.9 · 186.5 GB
Fits in less VRAM
Inkling Small
Thinking Machines
- 18.7 higher Score
- 55.2 GB less VRAM
Score 119.6 · 162.7 GB
Other vendor
Step 5 Preview
StepFun
- 27.1 higher Score
- 189.6 GB more VRAM
Score 128.0 · 407.5 GB
Similar models
10| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| DeepSeek V4 Flash 0731DeepSeek | 127.6 | 183 GB | —needs > 24 GB | similar score |
| GLM 5.3 FlashZ.ai | 124.1 | 205.2 GB | —needs > 24 GB | similar score |
| Beam 1.0 501B A23Breflection | 119.5 | 304 GB | —needs > 24 GB | similar score |
| DeepSeek V4 FlashDeepSeek | 115.3 | 175.1 GB | —needs > 24 GB | similar score |
| Qwen3.5 397B A17BAlibaba | 112.8 | 242.6 GB | —needs > 24 GB | similar score |
| MiniMax M3MiniMax | 112.6 | 258 GB | —needs > 24 GB | similar score |
| MiniMax M2.7MiniMax | 107.4 | 139.6 GB | —needs > 24 GB | similar score |
| DeepSeek V3.1 TerminusDeepSeek | 107.3 | 412.2 GB | —needs > 24 GB | similar score |
| MiniMax M2.1MiniMax | 106.2 | 139.6 GB | —needs > 24 GB | similar score |
| MiniMax M2MiniMax | 104.0 | 139.6 GB | —needs > 24 GB | similar score |
Specifications
- Publisher
- Z.ai
- Family
- GLM 4
- Parameters
- 358.3B
- Active parameters
- 35.6B per token
- Architecture
- Glm4MoeForCausalLM
- Context Length
- 131,072 tokens
- Vocabulary Size
- 151,552
- Release Date
- 2025-07-20
- License
- MIT
Frequently Asked Questions
How much VRAM does GLM 4.5 need?
GLM 4.5 requires 215.6 GB of VRAM at Q4_K_M, or 717.3 GB at BF16. Full 131K context adds up to 20.3 GB (235.9 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 358.3B × 4.8 bits ÷ 8 = 215 GB
KV Cache + Overhead ≈ 0.6 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 20.9 GB (at full 131K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run GLM 4.5?
No — GLM 4.5 requires at least 99.2 GB at IQ2_XXS, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
What's the best quantization for GLM 4.5?
For GLM 4.5, Q4_K_M (215.6 GB) offers the best balance of quality and VRAM usage. Q5_K_S (247.0 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 99.2 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run GLM 4.5 on a Mac?
Yes — Mac Pro M2 Ultra (128 GB) and 14 other Macs can run GLM 4.5, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 215.6 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio (M3 Ultra, 256GB).
Can I run GLM 4.5 locally?
Yes — GLM 4.5 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is GLM 4.5?
At Q4_K_M, GLM 4.5 can reach ~44 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B300 → 21.4 GB active ÷ (8000 × 0.65) = 4.12 ms, plus 92 layers × 0.055 ms = 5.06 ms, so 1000 ÷ 9.18 ms = ~109 tok/s
Estimated speed at Q4_K_M (215.6 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of GLM 4.5?
At Q4_K_M, the download is about 215.00 GB. The full-precision BF16 version is 716.68 GB. The smallest option (IQ2_XXS) is 98.54 GB.
Which GPUs can run GLM 4.5?
No single consumer GPU has enough VRAM to run GLM 4.5 at Q4_K_M (215.6 GB). Multi-GPU or professional hardware is required.
Which devices can run GLM 4.5?
2 devices with unified memory can run GLM 4.5 at Q4_K_M (215.6 GB), including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.