ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Q4_K_M: Smooth Fit 21.26 GB / 94.1 GB Usable

GLM-4-32B-0414 on NVIDIA H20 96GB (HBM3): VRAM Requirement & Speed Estimate

Yes. At Q4_K_M 4-bit quantization with a 16K context window, GLM-4-32B-0414 requires 21.26 GB of VRAM (19.4 GB weights + 1.02 GB KV cache + 0.84 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~144.6 tokens/s.

Concurrency (Batch):
Dense 32B·61L (d=6144)·48Q/2KV·Ctx 32K
Runnable Models →
GPUs
96 GB VRAM·4000 GB/s·Safe Usable 94.1 GB
Smooth Fit21.26 / 94.1 GB~144.6 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
21.26/ 94.1 GB usable

Physical 96 GB · Load 22.6%

Theoretical Decode Speed
~144.6tokens/s

Actual speed varies ~±25% by runtime

Max Context at Batch=1
32K/ 32K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 19.4 GBKV Cache (1×16K): 1.02 GBOverhead: 0.84 GB
🚀 Blazing Fast (Instant CoT)~3.5s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run glm4:32b

H20 96G Fits GLM-4-32B-0414 · Cost & TCO

Approx. $14,000 (~$146/GB VRAM) · 400W TDP

📊 All Quantization Levels for GLM-4-32B-0414 on H20 96G

Quantization Total VRAM (16K) Verdict Est. Speed Max Context
FP16 / BF16 66.75 GB Smooth Fit ~44.6 tok/s 32K
FP8 / Q8_0 36.15 GB Smooth Fit ~83.4 tok/s 32K
GGUF Q6_K 28.24 GB Smooth Fit ~107.7 tok/s 32K
GGUF Q5_K_M 24.69 GB Smooth Fit ~123.8 tok/s 32K
Q4_K_M / AWQ 21.26 GB Smooth Fit ~144.6 tok/s 32K
GGUF IQ4_XS 18.81 GB Smooth Fit ~164.5 tok/s 32K
GGUF Q3_K_M 17.43 GB Smooth Fit ~178.3 tok/s 32K
GGUF Q2_K 14.33 GB Smooth Fit ~219.7 tok/s 32K

💡 Recommended Alternative GPUs & Macs for GLM-4-32B-0414

Frequently Asked Questions

Can NVIDIA H20 96GB (HBM3) run GLM-4-32B-0414 locally?

Yes. At Q4_K_M 4-bit quantization with a 16K context window, GLM-4-32B-0414 requires 21.26 GB of VRAM (19.4 GB weights + 1.02 GB KV cache + 0.84 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~144.6 tokens/s.

What is the best quantization for GLM-4-32B-0414 on H20 96G?

The highest-quality quantization that fits cleanly on H20 96G is FP16 / BF16 (Lossless Full Precision), using 66.75 GB VRAM at ~44.6 tokens/s with up to 32K context.

How much VRAM does GLM-4-32B-0414 need in full FP16 precision?

In unquantized FP16/BF16 precision (16K context), GLM-4-32B-0414 requires 66.75 GB VRAM (64 GB for weights alone).