GLM-4-32B-0414 Architecture Parameters, Benchmarks & VRAM
Single-24GB-GPU friendly 32B coding model with ultra-compact 2-head GQA KV cache.
🧬 GLM-4-32B-0414 Architecture & Tensor Specifications
Verified directly against HuggingFace (zai-org/GLM-4-32B-0414) config.json & Model Card
1. Backbone & Dimensions
glm4- Total / Active Params
- 32B (Dense)
- Hidden Size (d_model)
- 6,144
- FFN Intermediate Size
- 23,040
- Activation / Norm
- SiLU (SwiGLU) · ε=1e-5
- Native Tensor Dtype
- BF16
2. Attention & Layer Layout
GQA- Num Hidden Layers
- 61 Layers
- Hidden Layer Layout
- 61-Layer Ultra-Lean GQA Dense Transformer (48 Q / 2 KV Heads, 24:1 Ratio)
- Attention Heads (Q / KV)
- 48 Q / 2 KV (24:1 GQA)
- Head Dimension
- 128
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 32,768 tokens
- RoPE Configuration
- RoPE (θ=10,000 · 原生 32,768 Context)
- Vocab Size
- 151,552 (Untied)
- Special Token IDs
- EOS: [151329,151336,151338] · PAD: 151329
🏆 Official Benchmark Results
Sourced from official System Card / Model Card & technical reports
Instruction-Following Evaluation (IFEval) benchmark for large language models, focusing on verifiable instructions with 25 types of instructions and around 500 prompts containing one or more verifiable constraints.
Berkeley Function Calling Leaderboard v3 (BFCL-v3) is an advanced benchmark that evaluates large language models' function calling capabilities through multi-turn and multi-step interactions.
Retail customer-service agent & API policy compliance
Multi-hop question answering across multiple documents
A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
💡 Pre-trained on 15T tokens; strong at HTML/SVG Artifacts and multi-turn tool calling.
⚡ GLM-4-32B-0414 Interactive VRAM & Token Speed Calculator
Physical 32 GB · Load 67.8%
Actual speed varies ~±25% by runtime
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run glm4:32bRTX 5090 Fits GLM-4-32B-0414 · Cost & TCO
Approx. $2,199 (~$69/GB VRAM) · 575W TDP
📊 GLM-4-32B-0414 Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 64 GB | 1.02 GB | 66.75 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 34 GB | 1.02 GB | 36.15 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 26.24 GB | 1.02 GB | 28.24 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 22.76 GB | 1.02 GB | 24.69 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 19.4 GB | 1.02 GB | 21.26 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 17 GB | 1.02 GB | 18.81 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 15.64 GB | 1.02 GB | 17.43 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 12.6 GB | 1.02 GB | 14.33 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for GLM-4-32B-0414 (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of GLM-4-32B-0414?
GLM-4-32B-0414 uses Glm4ForCausalLM with 61 layers (61-Layer Ultra-Lean GQA Dense Transformer (48 Q / 2 KV Heads, 24:1 Ratio)), hidden_size=6144, 48 Q heads / 2 KV heads (head_dim=128), and vocab_size=151,552.
How much VRAM does GLM-4-32B-0414 require?
At 4-bit quantization (Q4_K_M, 16K context), GLM-4-32B-0414 requires 21.26 GB VRAM. Full FP16 requires 66.75 GB.