GLM 5.3 Flash 320B-A18B
By Zhipu AI · China
Updated 2026-08-28
Overview
GLM 5.3 Flash from Zhipu — a multimodal MoE with 320B total/18B active params, 128k context, MIT-licensed, for multi-GPU server deployment of chat, code, and vision workloads.
When to pick this model
- Multi-GPU server deployments needing multimodal chat, code, and vision in one model
- Bilingual zh/en workloads at a lighter footprint than the full GLM 5.2 line
- Teams needing MIT licensing for multimodal flagship-class inference
- Long-form 128k-context tasks combining text and image input
VRAM requirements by quantization
| Quantization | VRAM required |
|---|---|
| Q4_K_M (recommended) | 186 GB |
| Q5_K_M | 227 GB |
| Q8_0 | 342 GB |
| FP16 (no quantization) | 640 GB |
VRAM figures include model weights plus a typical 8k KV cache and ~600 MB runtime overhead (Ollama / llama.cpp baseline). Add headroom for higher context lengths.
In practice, GLM 5.3 Flash 320B-A18B is server-class even at Q4_K_M (186 GB). Stepping up to Q8_0 nearly doubles the footprint to 342 GB, and unquantized FP16 weights take 640 GB — plan your GPU around the Q4 or Q5 figure unless you specifically need the higher fidelity.
Without a GPU, GLM 5.3 Flash 320B-A18B needs roughly 416 GB of system RAM to run on CPU via llama.cpp or Ollama — workable for background jobs, but far slower than GPU inference. Throughput estimates from our compatibility engine: around 9 tokens/sec on entry-level GPUs, on the order of 14 tokens/sec on a mid-range card, and up to 22 tokens/sec on high-end hardware — assuming the chosen quantization fully fits in VRAM.
What hardware do you need
The table below matches GLM 5.3 Flash 320B-A18B to common GPU memory tiers, using the highest-fidelity quantization that fully fits each card class. Spilling layers to system RAM works but costs most of the speed, so size your card to the quantization you actually want to run.
| GPU memory | Example cards | Best fit for GLM 5.3 Flash 320B-A18B |
|---|---|---|
| 8 GB | RTX 5070 Laptop, RTX 5060, RTX 5060 Ti 8GB | Does not fit — needs 186 GB at Q4_K_M |
| 12 GB | RTX 5070, RTX 5070 Ti Laptop, RTX 4080 Laptop | Does not fit — needs 186 GB at Q4_K_M |
| 16 GB | RTX 5080, RTX 4080 Super, Radeon RX 9070 XT | Does not fit — needs 186 GB at Q4_K_M |
| 24 GB | RTX 4090, Radeon RX 7900 XTX, RTX 5090 Laptop | Does not fit — needs 186 GB at Q4_K_M |
| 32 GB | RTX 5090 | Does not fit — needs 186 GB at Q4_K_M |
Which GPU should you buy to run GLM 5.3 Flash 320B-A18B?
To run GLM 5.3 Flash 320B-A18B locally at Q4, you need ~186 GB of VRAM. The best value for this is a Apple Mac Studio (64+ GB unified memory).
As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.
Strengths
- 320B total/18B active MoE — flagship multimodal capability
- Permissive MIT license
- Native 128k context
- Multimodal across chat, code, and vision
Limitations
- Still server-class multi-GPU hardware (~186GB VRAM at Q4)
- Weights gated on Hugging Face (requires acceptance)
- Throughput is moderate given the total parameter count
Typical workloads
In our catalog grid, GLM 5.3 Flash 320B-A18B is filed under Multi-GPU Server, Multimodal Code & Chat, Bilingual zh/en — the use cases where its size/quality trade-off makes the most sense. Its tags translate to concrete workloads: code generation and review (pair it with an editor integration like Continue.dev or Cline); vision-language work — screenshots, charts, scanned documents; multilingual workloads.
The 125k-token context window covers long chats and mid-sized documents, though very large retrieval workloads will need chunking. The MIT license is permissive, so shipping it inside a commercial product raises no special legal questions.
Architecture & training
Architecture: Multimodal MoE · 320B total parameters / 18B active per token · 128k context
Training: Flash variant of GLM 5.3 (Zhipu AI / THUDM), multimodal Mixture-of-Experts architecture. MIT license (GLM family).
A multimodal MoE flagship that trims total size versus GLM 5.2 while keeping chat, code, and vision in one MIT-licensed model.
Quick start
ollama pull glm-5.3-flashOr use the open-source MCP server to query this model from Claude Desktop, Cursor, or any MCP-compatible client.
Similar models worth comparing
Frequently asked questions
How much VRAM does GLM 5.3 Flash 320B-A18B need?
At the recommended Q4_K_M quantization, GLM 5.3 Flash 320B-A18B needs about 186 GB of VRAM. Q8_0 takes 342 GB, and unquantized FP16 weights take 640 GB.
Can GLM 5.3 Flash 320B-A18B run without a GPU?
Yes — with roughly 416 GB of system RAM it runs CPU-only through llama.cpp or Ollama. Expect a fraction of GPU speed, which is fine for background or batch jobs but slow for interactive chat.
What context window does GLM 5.3 Flash 320B-A18B support?
GLM 5.3 Flash 320B-A18B supports a 125k-token context window (128,000 tokens).
Can I use GLM 5.3 Flash 320B-A18B commercially?
Yes. GLM 5.3 Flash 320B-A18B is released under MIT, a permissive open-source license that allows commercial use, modification and redistribution.
How fast is GLM 5.3 Flash 320B-A18B on consumer hardware?
Our compatibility engine estimates on the order of 14 tokens/sec on a mid-range GPU and up to 22 tokens/sec on high-end cards, assuming the quantization fully fits in VRAM.
Which quantization of GLM 5.3 Flash 320B-A18B should I download first?
Start with Q4_K_M (186 GB) — the standard size/quality sweet spot. Step up to Q5_K_M or Q8_0 only if you have VRAM headroom. It does not fit a single 24 GB consumer card — plan for multi-GPU or server hardware.