GLM-5.3 — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
753B parameters in a 256-expert MoE (8 routed per token, about 40B active) with multi-head latent attention and DeepSeek-style sparse attention; 1M-token context. Every expert must be resident, so at Q4_K_M it needs a multi-GPU server. Z.ai serves it with SGLang and vLLM; Ollama lists only a cloud-hosted tag.
GLM-5.3 在 Q4_K_M 下约需 456 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)。包含 8K 上下文的 KV 缓存，因此数值高于上方的主数字。
| 量化 | 显存 | 显存 | 速度（估算） | 适配 |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 249.2 GB | — | 放不下 | |
| Q3_K_M 3.41 bpw | 322.6 GB | — | 放不下 | |
| Q4_K_M 4.83 bpw | 456.3 GB | — | 放不下 | |
| Q5_K_M 5.67 bpw | 535.4 GB | — | 放不下 | |
| Q6_K 6.56 bpw | 619.2 GB | — | 放不下 | |
| Q8_0 8.50 bpw | 801.9 GB | — | 放不下 | |
| F16 16.00 bpw | 1508.1 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 显存计算器 →
运行 GLM-5.3
GLM-5.3 is not practical on a single consumer GPU: it needs about 455.6 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
如何运行 GLM-5.3
No local Ollama tag. Ollama lists glm-5.3 only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine.
规格
Corroborated — Two or more independent sources agree on these figures, but the model card itself was not retrieved. Treat the numbers as good rather than confirmed. 尚未确认： active parameters.
- 参数量
- 753 Billion (40B active)
- 上下文窗口
- 1,048,576
- 架构
- Sparse MoE (256 experts, 8 active), MLA + sparse attention
- 提供商
- Zhipu AI (Z.ai)
- 许可证
- GLM-5.3 License (MIT-style, with a security-review clause for large Model-as-a-Service operators)
- Availability
- Weights downloadable
- Practical deployment
- Multi-GPU server
- 规格量化
- Q4_K_M
- 系统内存
- 512 GB
- 记录更新于
- 2026-10-13
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
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。
Independent evaluation — not vendor-claimed and not measured by us
| 基准 | 分数 | 来源 |
|---|---|---|
| AA Intelligence Index v4.3.2 (Max) | 45 | independent · Artificial Analysis · snapshot 2026-10-13 |
GLM-5.3 — 常见问题
How much VRAM does GLM-5.3 need?
About 456 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 run on an RTX 4090 (24 GB)?
No. GLM-5.3 needs about 456 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 locally?
No local Ollama tag. Ollama lists glm-5.3 only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine. Running the published tag would send your prompts to a hosted GPU rather than your own machine.