GLM-5.3-Flash (320B-A18B MoE) Architecture Parameters, Benchmarks & VRAM
Z.ai 320B total / 18B active MIT-licensed multimodal MoE optimized for agentic tool use and coding.
🧬 GLM-5.3-Flash (320B-A18B MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (zai-org/GLM-5.3-Flash) config.json & Model Card
1. Backbone & Dimensions
glm5_next- Total / Active Params
- 320B / 18B
- Hidden Size (d_model)
- 4,096
- FFN Intermediate Size
- 12,288
- MoE Expert Dim
- 2,048
- MoE Experts
- 288 Routed · Top-8 + 1 Shared
- Activation / Norm
- SiLU (SwiGLU) · ε=1e-5
- Native Tensor Dtype
- BF16
2. Attention & Layer Layout
MLA- Num Hidden Layers
- 45 Layers (11 Global KV)
- Hidden Layer Layout
- 45 Hybrid Layers: 34 Linear Attention + 11 DeepSeek Sparse Attention (DSA) + mHC
- Attention Heads (Q / KV)
- 64 Q / 64 KV
- Head Dimension
- 0
- Linear / KV Compression
- 34 zero-KV linear layers + 11 DSA MLA-compressed layers (kv_lora_rank=512)
- Vision Encoder / MTP
- ViT 24L · MTP 1L
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 1,048,576 tokens
- RoPE Configuration
- Hybrid DSA Position Encoding (原生 1,048,576 / 1M Context)
- Vocab Size
- 154,880 (Untied)
- Special Token IDs
- EOS: [154820,154827,154829] · PAD: 154820
📋 GLM-5.3-Flash (320B-A18B MoE) Official API & System Card Specifications
Verified against official Zhipu AI developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- glm-5.3-flash
- Release Date
- 2026-08-26
- Weight License
- Open-Weight (API Spec)
- Native Reasoning
- ✅ Supported
- Function Calling & JSON Schema
- ✅ 原生支持
- Batch API 异步批处理
- ✅ Supported (50% off)
2. Context, Output & Modalities
Limits- Max Context Window
- 256,000 tokens (256K)
- Max Output Tokens
- 131,000 tokens
- Input Modalities
- Text + Image + Video
- Output Modalities
- Text
- Direct China Access
- ✅ Direct CN + RMB
- Deployment Mode
- Open-Weight + Cloud API
3. Official Pricing & Cache
CNY / 1M- Input / 1M Tokens
- ¥ 0.8
- Cached Input / 1M
- ¥ 0.23 -71%
- Output / 1M Tokens
- ¥ 2.8
- Tiered Pricing Notes
- Zhipu BigModel: Official Zhipu BigModel API
🏆 Official Benchmark Results
Sourced from official System Card / Model Card & technical reports
PhD-level Google-proof science reasoning questions
Humanity's Last Exam evaluated with tool use
Terminal-Bench 2.1 is an updated release of the Terminal-Bench benchmark that tests AI agents' ability to operate a computer via the terminal.
DeepSWE is a software engineering agent benchmark evaluated with the mini-swe-agent harness, where each task is solved in an isolated container with no internet access. It measures an agent's ability to autonomously resolve real-world coding issues end to end.
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
💡 Supports reasoning_effort=low/high/max; pass clear_thinking=true in multi-turn chat.
📊 GLM-5.3-Flash (320B-A18B MoE) Full Benchmark Scorecard
15 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 3 items
Terminal-Bench 2.1 is an updated release of the Terminal-Bench benchmark that tests AI agents' ability to operate a computer via the terminal.
DeepSWE 1.1 evaluates software engineering agents using the mini-swe-agent harness.
NL2Repo evaluates long-horizon coding capabilities including repository-level understanding, where models must generate or modify code across entire repositories from natural language specifications.
🤖 Agents & Tools 5 items
Tool Decathlon is a comprehensive benchmark for evaluating AI agents' ability to use multiple tools across diverse task categories. It measures proficiency in tool selection, sequencing, and execution across ten different tool-use scenarios.
OfficeQA Pro evaluates AI models on professional knowledge-work questions and tasks drawn from real office workflows, including document analysis, spreadsheet reasoning, and information synthesis across business domains.
GDPval-AA evaluates AI agents on economically valuable professional knowledge-work tasks and reports performance as an Elo score.
AutomationBench is a tool-use benchmark that evaluates AI agents on automating real-world workflows, testing their ability to orchestrate tools and complete multi-step automation tasks.
Agents' Last Exam is a challenging benchmark for AI agents on hard, long-horizon tasks that test sustained reasoning, planning, and tool use, reported with and without tool access.
👁️ Multimodal 5 items
CharXiv-R is the reasoning component of the CharXiv benchmark, focusing on complex reasoning questions that require synthesizing information across visual chart elements.
MMVU (Multimodal Multi-disciplinary Video Understanding) is a benchmark for evaluating multimodal models on video understanding tasks across multiple disciplines, testing comprehension and reasoning capabilities on video content.
Chartography evaluates professional chart understanding with 100 expert-written tasks drawn from real charts across science, engineering, finance, manufacturing, supply chain, and healthcare.
A comprehensive multi-modal video understanding benchmark covering 20 challenging video tasks that require temporal understanding beyond single-frame analysis.
A benchmark for early-stage visual reasoning and perception on child-like vision tasks.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ GLM-5.3-Flash (320B-A18B MoE) Interactive VRAM & Token Speed Calculator
Physical 512 GB · Load 43.3%
MoE: reads only 10.91 GB active weights/step
MLA compresses multi-batch KV footprint
OLLAMA_CONTEXT_LENGTH=16384 ollama run glm5:flashMac Studio M5 Ultra 512G Fits GLM-5.3-Flash (320B-A18B MoE) · Cost & TCO
Approx. $9,499 (~$19/GB VRAM) · 190W TDP
📊 GLM-5.3-Flash (320B-A18B MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 640 GB | 0.21 GB | 654.31 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 340 GB | 0.21 GB | 348.31 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 262.4 GB | 0.21 GB | 269.16 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 227.6 GB | 0.21 GB | 233.66 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 194 GB | 0.21 GB | 199.39 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 170 GB | 0.21 GB | 174.91 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 156.4 GB | 0.21 GB | 161.04 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 126 GB | 0.21 GB | 130.03 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for GLM-5.3-Flash (320B-A18B MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of GLM-5.3-Flash (320B-A18B MoE)?
GLM-5.3-Flash (320B-A18B MoE) uses Glm5NextForConditionalGeneration with 45 layers (45 Hybrid Layers: 34 Linear Attention + 11 DeepSeek Sparse Attention (DSA) + mHC), hidden_size=4096, 64 Q heads / 64 KV heads (head_dim=0), and vocab_size=154,880.
How much VRAM does GLM-5.3-Flash (320B-A18B MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), GLM-5.3-Flash (320B-A18B MoE) requires 199.39 GB VRAM. Full FP16 requires 654.31 GB.