GLM-4 · MIT · HF: zai-org/GLM-4-32B-0414
GLM-4 32B 0414 — VRAM, KV Cache & Hardware Matrix
Serverless API Rate
$0.10 in / $0.30 out per 1M
Parameters
32.3B Dense
Attention
GQA (2 KV / 48 Q)
Layers × Hidden
61 × 6144
Max Native Context
32k tokens
MoE Experts
N/A (Dense)
Quantization Memory Footprint
| Quantization | BPW | Weights Only | Total @ 8k Ctx | Total @ 32k Ctx | Quality |
|---|---|---|---|---|---|
| FP16 | 16.00 | 61.37 GiB | 63.44 GiB | 64.87 GiB | Lossless Reference |
| Q8_0 | 8.50 | 32.60 GiB | 34.67 GiB | 36.10 GiB | Near-Lossless |
| Q6_K | 6.57 | 25.20 GiB | 27.27 GiB | 28.70 GiB | High Precision |
| Q5_K_M | 5.69 | 21.82 GiB | 23.89 GiB | 25.32 GiB | Balanced High |
| Q4_K_M | 4.85 | 18.60 GiB | 20.67 GiB | 22.10 GiB | Recommended 4-bit |
| MLX_4BIT | 4.50 | 17.26 GiB | 19.33 GiB | 20.76 GiB | Apple Silicon Optimized |
| IQ4_XS | 4.25 | 16.30 GiB | 18.37 GiB | 19.80 GiB | Compact 4-bit |
| Q3_K_M | 3.91 | 15.00 GiB | 17.07 GiB | 18.50 GiB | Aggressive Compression |
Hardware Compatibility & Speed Leaderboard (Q4_K_M @ 32k Context)
| Hardware | Status | Total / Limit | Decode | TTFT (4k) | Max Safe Ctx | Inspect |
|---|---|---|---|---|---|---|
| Cloud NVIDIA H200 SXM (141 GB) | COMFORTABLE | 22.1 / 140.4 GiB | 183.1 tok/s | 0.42s | 32k | Simulate → |
| Cloud NVIDIA H100 SXM (80 GB) | COMFORTABLE | 22.1 / 79.4 GiB | 127.8 tok/s | 0.42s | 32k | Simulate → |
| Cloud NVIDIA A100 SXM (80 GB) | COMFORTABLE | 22.1 / 79.4 GiB | 75.9 tok/s | 1.34s | 32k | Simulate → |
| NVIDIA RTX PRO 6000 Blackwell (96 GB) | COMFORTABLE | 22.1 / 95.0 GiB | 68.4 tok/s | 1.67s | 32k | Simulate → |
| NVIDIA RTX 5090 (32 GB) | COMFORTABLE | 22.1 / 31.0 GiB | 66.7 tok/s | 2.07s | 32k | Simulate → |
| Cloud NVIDIA L40S (48 GB) | COMFORTABLE | 22.1 / 47.4 GiB | 32.2 tok/s | 1.20s | 32k | Simulate → |
| Apple Mac Studio M3 Ultra (256 GB) | COMFORTABLE | 21.9 / 199.7 GiB | 29.0 tok/s | 8.28s | 32k | Simulate → |
| Apple Mac Studio M3 Ultra (512 GB) | COMFORTABLE | 21.9 / 399.4 GiB | 29.0 tok/s | 8.28s | 32k | Simulate → |
| 2x NVIDIA RTX 3090 Rig (48 GB) | COMFORTABLE | 22.1 / 46.6 GiB | 28.6 tok/s | 7.78s | 32k | Simulate → |
| Apple Mac Studio M2 Ultra (192 GB) | COMFORTABLE | 21.9 / 149.8 GiB | 28.3 tok/s | 9.13s | 32k | Simulate → |
| Apple Mac Studio M4 Max (64 GB) | COMFORTABLE | 21.9 / 49.9 GiB | 19.1 tok/s | 12.59s | 32k | Simulate → |
| Apple Mac Studio M4 Max (128 GB) | COMFORTABLE | 21.9 / 99.8 GiB | 19.1 tok/s | 12.59s | 32k | Simulate → |
| Apple MacBook Pro M3 Max (64 GB) | COMFORTABLE | 21.9 / 49.9 GiB | 13.8 tok/s | 16.86s | 32k | Simulate → |
| Apple MacBook Pro M3 Max (128 GB) | COMFORTABLE | 21.9 / 99.8 GiB | 13.8 tok/s | 16.86s | 32k | Simulate → |
| Apple Mac Mini M4 Pro (48 GB) | COMFORTABLE | 21.9 / 37.4 GiB | 9.2 tok/s | 26.01s | 32k | Simulate → |
| Apple Mac Mini M4 Pro (64 GB) | COMFORTABLE | 21.9 / 49.9 GiB | 9.2 tok/s | 26.01s | 32k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (128 GB) | COMFORTABLE | 21.9 / 96.0 GiB | 8.6 tok/s | 9.37s | 32k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (64 GB) | COMFORTABLE | 21.9 / 48.0 GiB | 8.6 tok/s | 9.37s | 32k | Simulate → |
| Ryzen AI 9 HX 370 Mini-PC (64 GB) | COMFORTABLE | 21.9 / 48.0 GiB | 3.8 tok/s | 26.82s | 32k | Simulate → |
| NVIDIA RTX 4090 (24 GB) | TIGHT | 22.1 / 23.0 GiB | 37.5 tok/s | 2.62s | 32k | Simulate → |
| NVIDIA RTX 3090 (24 GB Used) | TIGHT | 22.1 / 23.0 GiB | 34.0 tok/s | 6.33s | 32k | Simulate → |
| AMD Radeon RX 7900 XTX (24 GB) | TIGHT | 22.1 / 23.0 GiB | 33.1 tok/s | 4.38s | 32k | Simulate → |
| NVIDIA RTX 5080 (16 GB) | OOM | 22.1 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5070 Ti (16 GB) | OOM | 22.1 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5060 Ti (16 GB) | OOM | 22.1 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3060 (12 GB) | OOM | 22.1 / 11.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (16 GB) | OOM | 21.9 / 10.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (24 GB) | OOM | 21.9 / 18.0 GiB | OOM | — | 2k | Simulate → |
Interactive GLM-4 32B 0414 Simulator
32.3B Dense · GQA
1792 GB/s · 250 TFLOPS
3. Weight Quantization
4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 32k
Memory Allocation Footprint 22.10 / 95.00 GiB usable
Weights: 18.60 GiB
KV Cache: 1.91 GiB
Overhead: 1.59 GiB
Headroom: 72.90 GiB
Decode Speed
68.4
tok/s (batch=1)
Prefill Speed
2512
prompt tok/s
TTFT Latency
1.67s
for 4,096 tok
Active Params
32.3B
of 32.3B total
llama.cpp Server Command
llama-server -hf zai-org/GLM-4-32B-0414:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080| Context | KV Cache | Total VRAM | Decode | Status |
|---|---|---|---|---|
| 4k | 0.24 GiB | 20.01 GiB | 74.4 tok/s | COMFORTABLE |
| 8k | 0.48 GiB | 20.67 GiB | 73.5 tok/s | COMFORTABLE |
| 16k | 0.95 GiB | 21.15 GiB | 71.7 tok/s | COMFORTABLE |
| 32k | 1.91 GiB | 22.10 GiB | 68.4 tok/s | COMFORTABLE |
Buy-vs-Cloud-vs-API Monthly TCO Break-Even
USE SERVERLESS APIOwn Hardware/mo
$166.01
66.5h active + power
Cloud GPU/mo
$141.44
@$1.85/hr spot/pod
Serverless API/mo
$9.50
API cheaper
Recommended Hardware & Cloud Deployment
Contextual to fit statusEngineering FAQ — GLM-4 32B 0414
How much VRAM does GLM-4 32B 0414 require in Q4_K_M?
GLM-4 32B 0414 weights occupy 18.60 GiB in Q4_K_M. Including 32k FP16 KV cache and runtime buffers, total memory is 22.10 GiB.
What is the cheapest local hardware that runs GLM-4 32B 0414?
NVIDIA RTX PRO 6000 Blackwell (96 GB) (96 GB) runs GLM-4 32B 0414 Q4_K_M at ~68.4 tok/s decode and 1.67s TTFT.
How does GLM-4 32B 0414 scale KV cache at long context?
GLM-4 32B 0414 uses GQA with 2 KV heads across 61 layers (128 head_dim).