GLM-5.3-Flash 320B-A18B — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
320 billion parameters must all be resident — about 194 GB at Q4_K_M, so this is a multi-GPU or large-unified-memory model, not a single-card one. Only 18B activate per token, which is why it decodes at roughly the speed of a 18B dense model once loaded. Natively multimodal across text, image and video with a 1,048,576-token context. MIT licensed. Ollama publishes only a `:cloud` tag; local runners use the community GGUF conversions.
GLM-5.3-Flash 320B-A18B 在 Q4_K_M 下约需 194 GB 显存——量化权重加框架开销，不含 KV 缓存。在 Apple Silicon 上，这部分来自统一内存。
按量化级别的显存与速度
计算基准：NVIDIA RTX 4090 (24 GB)。仅含权重与开销：该模型架构未公开，因此未计入 KV 缓存。
| 量化 | 显存 | 显存 | 速度（估算） | 适配 |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 106 GB | — | 放不下 | |
| Q3_K_M 3.41 bpw | 137.2 GB | — | 放不下 | |
| Q4_K_M 4.83 bpw | 194 GB | — | 放不下 | |
| Q5_K_M 5.67 bpw | 227.6 GB | — | 放不下 | |
| Q6_K 6.56 bpw | 263.2 GB | — | 放不下 | |
| Q8_0 8.50 bpw | 340.8 GB | — | 放不下 | |
| F16 16.00 bpw | 640.8 GB | — | 放不下 |
黑色标记 = NVIDIA RTX 4090 (24 GB) 上的可用显存。 估算来自内存带宽屋顶线模型，详见 方法说明页. GLM-5.3-Flash 320B-A18B 显存计算器 →
运行 GLM-5.3-Flash 320B-A18B
目录中能运行 GLM-5.3-Flash 320B-A18B 的最便宜 GPU 是 Apple M3 Ultra (512 GB).
作为亚马逊联盟成员，我们从符合条件的购买中获得收入。云 GPU 链接为推荐链接——我们可能获得佣金，您无需额外付费。
如何运行 GLM-5.3-Flash 320B-A18B
The Ollama tag `glm-5.3-flash:cloud` is cloud-hosted, so there is no local run command for this model.
规格
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- 参数量
- 320 Billion (18B active)
- 上下文窗口
- 1,048,576
- 架构
- Hybrid-Attention MoE (multimodal)
- 提供商
- Zhipu AI (Z.ai)
- 许可证
- MIT
- Practical deployment
- Workstation (large unified memory or pro GPU)
- 规格量化
- Q4_K_M
- 系统内存
- 256 GB
- 记录更新于
- 2026-09-11
Commercial use permitted. No usage restrictions beyond attribution.
另请参阅 GLM-5.3
质量与使用场景
评分由模型作者或独立评测方发布——衡量质量而非吞吐量，并非我们实测。
| 基准 | 分数 | 来源 |
|---|---|---|
| DeepSWE | 63.4 / 100 % | 厂商声称 · https://www.marktechpost.com/2026/08/26/z-ai-releases-glm-5-3-flash-a-320b-a18b-natively-multimodal-moe-with-a-1m-token-context/ |
Independent evaluation — not vendor-claimed and not measured by us
| 基准 | 分数 | 来源 |
|---|---|---|
| AA Intelligence Index v4.3.2 | 42 | independent · Artificial Analysis · snapshot 2026-10-13 |
我的 GPU 能运行 GLM-5.3-Flash 320B-A18B 吗？
- GLM-5.3-Flash 在 AMD Ryzen AI Max+ 395 上
- GLM-5.3-Flash 在 Apple M1 Ultra 上
- GLM-5.3-Flash 在 Apple M2 Max 上
- GLM-5.3-Flash 在 Apple M2 Ultra 上
- GLM-5.3-Flash 在 Apple M4 Max 上
- GLM-5.3-Flash 在 Apple M5 Max 上
- GLM-5.3-Flash 在 NVIDIA A100 80GB (PCIe) 上
- GLM-5.3-Flash 在 NVIDIA DGX Spark 上
- GLM-5.3-Flash 在 NVIDIA H100 80GB (PCIe) 上
- GLM-5.3-Flash 在 NVIDIA RTX PRO 6000 Blackwell 上
GLM-5.3-Flash 320B-A18B — 常见问题
How much VRAM does GLM-5.3-Flash 320B-A18B need?
About 194 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does GLM-5.3-Flash 320B-A18B run on an RTX 4090 (24 GB)?
No. GLM-5.3-Flash 320B-A18B needs about 194 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run GLM-5.3-Flash 320B-A18B locally?
The Ollama tag `glm-5.3-flash:cloud` is cloud-hosted, so there is no local run command for this model. Running the published tag would send your prompts to a hosted GPU rather than your own machine.