GLM-5.2 (753B MoE) Architecture Parameters, Benchmarks & VRAM
MIT-licensed 753B 1M-context flagship MoE from Zhipu AI scoring 91.2% on GPQA Diamond.
🧬 GLM-5.2 (753B MoE) Architecture & Tensor Specifications
Verified directly against HuggingFace (zai-org/GLM-5.2) config.json & Model Card
1. Backbone & Dimensions
glm_moe_dsa- Total / Active Params
- 753B / 40B
- Hidden Size (d_model)
- 6,144
- FFN Intermediate Size
- 12,288
- MoE Expert Dim
- 2,048
- MoE Experts
- 256 Routed · Top-8 + 1 Shared
- Activation / Norm
- SiLU (SwiGLU) · ε=0.00001
- Native Tensor Dtype
- BF16 / FP8
2. Attention & Layer Layout
MLA- Num Hidden Layers
- 78 Layers
- Hidden Layer Layout
- 78-Layer Sparse MoE (256 Experts, Top-8 Active)
- Attention Heads (Q / KV)
- 64 Q / 64 KV
- Head Dimension
- 192
3. Context, RoPE & Tokenizer
Tokenizer- Native Context Length
- 1,048,576 tokens
- RoPE Configuration
- RoPE (原生 1,048,576 Context)
- Vocab Size
- 154,880 (Untied)
📋 GLM-5.2 (753B MoE) Official API & System Card Specifications
Verified against official Zhipu AI developer documentation & pricing
1. Identity & Capabilities
API Spec- API Model ID
- glm-5.2
- Release Date
- 2026-06-13
- 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
- 1,000,000 tokens (1000K)
- Max Output Tokens
- 131,000 tokens
- Input Modalities
- Text
- Output Modalities
- Text
- Direct China Access
- ✅ Direct CN + RMB
- Deployment Mode
- Open-Weight + Cloud API
3. Official Pricing & Cache
CNY / 1M- Input / 1M Tokens
- ¥ 8
- Cached Input / 1M
- ¥ 2 -75%
- Output / 1M Tokens
- ¥ 28
- Tiered Pricing Notes
- Zhipu BigModel: Official Zhipu BigModel API
🎛️ Official Sampling & Reasoning Parameters
Recommended generation settings & API parameter constraints
📊 GLM-5.2 (753B MoE) Full Benchmark Scorecard
19 public benchmarks · ranks among 154 models · synced 2026-11-20
💻 Coding 11 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.
MCP Atlas is a benchmark for evaluating AI models on scaled tool use capabilities, measuring how well models can coordinate and utilize multiple tools across complex multi-step tasks.
FrontierSWE measures whether an agent can complete open-ended technical projects at the scale of hours to tens of hours, spanning systems optimization, large-scale code construction, and applied ML research. Performance is reported as a dominance score, where higher is better.
Program Bench evaluates code-generation agents by asking them to recreate a program's behavior from only a compiled binary and documentation.
SWE-Bench Pro is an advanced version of SWE-Bench that evaluates language models on complex, real-world software engineering tasks requiring extended reasoning and multi-step problem solving.
NL2Repo evaluates long-horizon coding capabilities including repository-level understanding, where models must generate or modify code across entire repositories from natural language specifications.
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.
DeepSWE 1.1 evaluates software engineering agents using the mini-swe-agent harness.
PostTrainBench evaluates a model's ability to autonomously post-train base models. Given pretrain-only base models, the agent must complete the full pipeline of data synthesis, training, evaluation, and iteration within a time budget, scored across downstream benchmarks such as…
FrontierCode 1.1 evaluates whether coding-agent changes are mergeable, using unit tests, maintainer-defined rubrics, and verifiers. Runs flagged for unfair internet use receive a zero score.
SWE-Marathon is an ultra-long-horizon software engineering benchmark covering tasks such as building compilers, optimizing kernels, and developing production-grade services. It measures whether agents can sustain quality across extremely long engineering trajectories.
📐 Math 6 items
All 30 problems from the 2026 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999.
Harvard-MIT Mathematics Tournament 2025 - A prestigious student-organized mathematics competition for high school students featuring two tournaments (November 2025 at MIT and February 2026 at Harvard) with individual tests, team rounds, and guts rounds.
HMMT February 2026 is a math competition benchmark based on problems from the Harvard-MIT Mathematics Tournament, testing advanced mathematical problem-solving and reasoning.
IMO-AnswerBench is a benchmark for evaluating mathematical reasoning capabilities on International Mathematical Olympiad (IMO) problems, focusing on answer generation and verification.
Humanity's Last Exam (HLE) is a multi-modal academic benchmark with 2,500 questions across mathematics, humanities, and natural sciences, designed to test LLM capabilities at the frontier of human knowledge with unambiguous, verifiable solutions.
CritPT is a challenging reasoning benchmark reported by Qwen for evaluating frontier mathematical and critical problem-solving capability.
Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).
⚡ GLM-5.2 (753B MoE) Interactive VRAM & Token Speed Calculator
Physical 96 GB · Load 497%
Exceeds usable VRAM by 373.48 GB
MLA compresses multi-batch KV footprint
OLLAMA_CONTEXT_LENGTH=16384 ollama run glm-5-2-744bInsufficient Local VRAM? Ways to Run GLM-5.2 (753B MoE)
Rent an hourly high-VRAM cloud GPU pod, or check the 12-month Buy vs Cloud vs API break-even curve.
📊 GLM-5.2 (753B MoE) Quantization VRAM Breakdown (16K Context)
| Quantization | Bits/Weight | Weights | KV Cache (16K) | Total VRAM | Quality Note |
|---|---|---|---|---|---|
| FP16 / BF16 | 16 bpw | 1506 GB | 1.47 GB | 1538.04 GB | Lossless Full Precision |
| FP8 / Q8_0 | 8.5 bpw | 800.06 GB | 1.47 GB | 817.99 GB | 99.9% Near-Lossless |
| GGUF Q6_K | 6.56 bpw | 617.46 GB | 1.47 GB | 631.73 GB | Ultra High Quality |
| GGUF Q5_K_M | 5.69 bpw | 535.57 GB | 1.47 GB | 548.2 GB | High Balanced Quality |
| Q4_K_M / AWQ | 4.85 bpw | 456.51 GB | 1.47 GB | 467.56 GB | Sweet Spot (Recommended) |
| GGUF IQ4_XS | 4.25 bpw | 400.03 GB | 1.47 GB | 409.95 GB | Compact 4-bit |
| GGUF Q3_K_M | 3.91 bpw | 368.03 GB | 1.47 GB | 377.31 GB | Moderate Perplexity Loss |
| GGUF Q2_K | 3.15 bpw | 296.49 GB | 1.47 GB | 304.35 GB | Extreme Compression |
🖥️ GPU & Apple Silicon Compatibility for GLM-5.2 (753B MoE) (Q4_K_M)
Frequently Asked Questions
What are the exact architecture parameters of GLM-5.2 (753B MoE)?
GLM-5.2 (753B MoE) uses GlmMoeDsaForCausalLM with 78 layers (78-Layer Sparse MoE (256 Experts, Top-8 Active)), hidden_size=6144, 64 Q heads / 64 KV heads (head_dim=192), and vocab_size=154,880.
How much VRAM does GLM-5.2 (753B MoE) require?
At 4-bit quantization (Q4_K_M, 16K context), GLM-5.2 (753B MoE) requires 467.56 GB VRAM. Full FP16 requires 1538.04 GB.