The first natively multimodal model in the GLM-5 series. Featuring 320B total parameters with only 18B active parameters, GLM-5.3-Flash delivers Claude Opus 4.8 coding & agentic intelligence at $0.045 per task (1/10th the cost), powered by a hybrid Linear-Sparse attention engine and served at scale on domestic AI chips.
This repository serves as the definitive reference guide and benchmark suite for GLM-5.3-Flash—Z.ai's state-of-the-art 320B Mixture-of-Experts (MoE) multimodal model.
- Core Model Features: 320B total parameters with 18B active parameters across 45 layers; 1M token context window powered by IndexPool 4:1 key compression; hybrid Linear-Sparse attention architecture; natively trained on 30T tokens with visual self-verification; 10x lower task cost ($0.045/task) compared to frontier competitors like Claude Opus 4.8.
- Benchmark Performance: Achieves top-tier scores across software engineering (DeepSWE 63.4), agentic automation (AutomationBench 48.8), terminal coding (TerminalBench 84.3), tool usage (Toolathlon 78.4), and agentic judgment (GDPval-AA 1773 Elo).
- Battle-Testing Suite: Includes 3 curated real-world evaluation scenarios designed to stress-test multimodal code generation (Single-File UI from Dashboard Screenshots), data reasoning (Data-to-Python Report & Charting), and long-context compliance (Massive Apple 10-K Financial Document Extraction to Styled HTML).
Before its official release, Z.ai anonymously deployed GLM-5.3-Flash under the mystery alias ox-alpha on OpenCode and OpenRouter.
- 🕵️ Viral Developer Mystery: Within days,
ox-alphabecame the #1 trending model of the week. Developers logged over 250M+ continuous token runs on complex multi-hour agent workflows. - ⚡ Domestic Accelerator Scale: All
ox-alphatraffic was served natively across tens of thousands of domestic Chinese AI accelerators. - 🤖 Self-Optimizing Infrastructure: The serving stack was tuned by GLM-5.3 infrastructure agents running inside the cluster to optimize kernels and scheduling.
flowchart TD
A[30T Multimodal Data Stream] --> B[SigLIP2 / Native Visual Patching]
B --> C[IndexPool: 4:1 Key Compression]
C --> D[Hybrid Attention Engine]
D --> E[Linear Attention: Local State]
D --> F[Sparse Attention: Global Retrieval]
E & F --> G[mHC Hyper-Connections]
G --> H[Disaggregated EPD Serving Engine]
H --> I[Output: Code, GUI Actions & Reasoning]
| Parameter / Feature | GLM-5.3-Flash Specification | Performance & Cost Advantage |
|---|---|---|
| Total Parameters | 320 Billion |
Full frontier model knowledge base |
| Active Parameters | 18 Billion (MoE) |
Ultra-low compute per token generated |
| Active Layers | 45 Layers |
Halved layer count vs GLM-4.5 (92 layers) |
| Context Window | 1,000,000 Tokens (1M) |
Native long-context with lightweight IndexPool |
| Task Cost | $0.045 / task |
10x cheaper than Claude Opus 4.8 ($0.50) |
| Attention Compute | 3.0x Lower |
Lowest attention compute among all 300B+ MoEs |
| KV Cache Footprint | 4.4x Reduction |
Weighted key vector pooling via IndexPool |
| Training Data | 30 Trillion Tokens |
Multimodal corpus with visual self-verification |
GLM-5.3-Flash pushes the Artificial Analysis Pareto frontier, matching or outperforming Claude Opus 4.8 and GPT-5.6 Terra on key coding and agentic benchmarks:
| Benchmark Category | Benchmark Name | GLM-5.3-Flash | GLM-5.2 | Claude Opus 4.8 | DeepSeek-V4-Vision | Gemini 3.7 Flash |
|---|---|---|---|---|---|---|
| Software Engineering | DeepSWE v1.1 🥇 | 63.4 | 46.2 | 58.0 | 59.3 | 65.3 |
| Agentic Automation | AutomationBench 🥇 | 48.8 | 26.2 | 41.0 | 38.8 | 52.3 |
| Terminal Coding | TerminalBench 2.1 | 84.3 | 81.0 | 85.0 | 83.9 | 85.8 |
| Tool Execution | Toolathlon Verified 🥇 | 78.4 | 59.9 | 76.2 | 75.9 | - |
| Agentic Judgment | GDPval-AA v2 🥇 | 1773 Elo | 1504 | 1582 | 1675 | 1527 |
| Visual Reasoning | CharXiv w/ Tools | 89.4 | - | 89.9 | 80.4 | 88.7 |
| Chart Analysis | Chartography | 78.0 | - | 75.0 | 64.3 | 65.0 |
Use these 3 battle-testing scenarios to evaluate GLM-5.3-Flash against competitor models:
Tests visual hierarchy extraction from screenshots into clean responsive HTML without JS frameworks.
| Field | Description |
|---|---|
| Attachment | Screenshot of a complex open-source admin dashboard (Tabler Admin Dashboard). |
| Prompt | "Based on this single screenshot of an admin dashboard, recreate the layout in a single HTML file. You must use only HTML and inline CSS (or a <style> block). Do not use any JavaScript, Bootstrap, or external libraries. Ensure the grid system, sidebar navigation, and typography accurately reflect the visual hierarchy of the image. Output the final result as a single index.html code block." |
| Why Competitors Fail | Hallucinate non-existent Bootstrap classes or rely on JavaScript for grid alignment. Struggle to keep self-contained CSS coherent. |
Tests multi-modal reasoning, data cleaning, and executable code generation without execution bugs.
| Field | Description |
|---|---|
| Attachment | Large raw CSV dataset (UCI Bank Marketing Dataset). |
| Prompt | "Analyze this dataset regarding marketing campaign results. First, output a single Markdown file (report.md) detailing the key trends and correlations. Second, write a single, self-contained Python script (analysis.py) using pandas and matplotlib that reads this CSV, cleans the data, and generates three specific charts illustrating the trends you found. The Python script must run perfectly on the raw file without any edits." |
| Why Competitors Fail | Reference incorrect column names, crash on nulls/missing types, or write reports that contradict the script's output. |
Tests 1M-token context retrieval and strict semantic HTML output compliance on long corporate filings.
| Field | Description |
|---|---|
| Attachment | Massive financial report PDF/text (Apple Inc. 10-K Annual Report SEC EDGAR). |
| Prompt | "Read this massive financial document. Extract the core revenue metrics, risk factors, and forward-looking statements. Output your findings as a highly styled, single-file HTML executive summary (summary.html). Use semantic HTML elements and HTML tables for the financial data. For every metric or risk factor, you must include a citation specifying the exact section from the source document." |
| Why Competitors Fail | Truncate long context, drop crucial risk factors, hallucinate financial numbers, or emit broken unclosed HTML table tags. |
GLM-5.3-Flash weights are open under MIT license on HuggingFace.
| Quantization / Format | Model Size | VRAM / RAM Target | Hardware Recommendation |
|---|---|---|---|
| NVFP4 (NVIDIA FP4) | ~160 GB | 256 GB | 1x Mac Studio M5 Ultra (256GB) or 2x DGX Spark |
| UD-Q4_K_XL | ~175 GB | 256 GB RAM + 16GB VRAM | Dual Xeon / Epyc Workstation (8-Channel DDR4/DDR5) |
| FP8 (Original Weights) | ~320 GB | 384 GB VRAM | 8x RTX 3090 / 4090 or Cluster Setup |
# Install SGLang for GLM-5.3-Flash
pip install --upgrade sglang
# Launch inference server with disaggregated EPD support
python3 -m sglang.launch_server \
--model-path zai-org/GLM-5.3-Flash \
--tp 8 \
--trust-remote-code \
--port 30000.
├── README.md # Main documentation & 3 battle-testing prompts
└── assets/
├── glm_5_3_flash_overview.html # Interactive editorial infographic HTML dashboard
└── glm_5_3_flash_overview.md # Synchronized Markdown report
- 🌐 Z.ai Official Platform: z.ai
- 📝 Technical Blog Post: z.ai/blog/glm-5.3-flash
- 🤗 HuggingFace Weights: zai-org/GLM-5.3-Flash
- 🚀 OpenRouter Model: OpenRouter ox-alpha / GLM-5.3-Flash
GLM-5.3-Flash Z.ai ox-alpha 320B MoE 18B Active 1M Context IndexPool Linear-Sparse Attention Claude Opus 4.8 Benchmark DeepSWE AutomationBench TerminalBench Toolathlon GDPval-AA SGLang vLLM NVFP4 Domestic AI Chips OpenRouter HuggingFace