GLM 5.3 on Ollama: Cloud Model, Local Options and Hardware Needs

Ollama lists GLM 5.3 as glm-5.3:cloud, a hosted model with a 1M-token context. There is no local Ollama tag, but GGUF quants from 217 GB run with llama.cpp. Commands, sizes and limits.

GLM 5.3 on Ollama: Cloud Model, Local Options and Hardware Needs

Quick answer: Ollama lists GLM 5.3 as glm-5.3:cloud. It runs on Ollama's cloud, not on your machine: ollama run glm-5.3:cloud. The Ollama library has no local tag for GLM 5.3. To run it on your own hardware, use the open weights Z.AI published on Hugging Face on August 25, 2026, or a GGUF quant such as unsloth/GLM-5.3-GGUF. Even the smallest of those quants is about 217 GB.


What Ollama offers today

ItemValue
Model tagglm-5.3:cloud (the only tag)
Where it runsOllama's cloud, hosted in the United States and Europe
Context window1M tokens
InputText only
Size753B parameters
Price on Ollama$1.40 input, $0.26 cached input, $4.40 output per 1M tokens
Reasoningreasoning_effort set to low, high or max (default max)

Ollama's model page says the cloud model follows its privacy policy with zero data retention. GLM 5.2 and GLM 5.3 Flash are also listed as cloud-only tags (glm-5.2:cloud, glm-5.3-flash:cloud).

Run GLM 5.3 through Ollama

You need an Ollama account and an API key for cloud models.

# Chat in the terminal
ollama run glm-5.3:cloud

# Start a coding agent on GLM 5.3
ollama launch claude --model glm-5.3:cloud
ollama launch opencode --model glm-5.3:cloud

Call it over HTTP:

curl https://ollama.com/api/chat \
  -H "Authorization: Bearer $OLLAMA_API_KEY" \
  -d '{
    "model": "glm-5.3",
    "messages": [{"role": "user", "content": "Write a quicksort in Python"}]
  }'

Thinking is always on in GLM 5.3. Expect reasoning tokens in long tasks, and set reasoning_effort to low when you want shorter answers.

Running GLM 5.3 on your own hardware

Z.AI released the weights on August 25, 2026 as zai-org/GLM-5.3 (FP8) and zai-org/GLM-5.3-BF16, under a custom glm-5.3 license. Unsloth published GGUF quants on August 28. Sizes below are the totals of each quant folder on Hugging Face:

FormatDownload sizeTypical setup
UD-IQ1_S (1-bit dynamic)~217 GB256 GB of RAM or unified memory, slow
UD-Q2_K_XL~254 GB256 GB+ plus a GPU for offload
UD-Q4_K_XL~467 GBMulti-GPU server or 512 GB of memory
Q8_0~801 GB8× 80 GB GPUs or more
FP8 (official)~753 GB8× H100/H200 with SGLang or vLLM

Each GGUF quant is split into several files. Load them with llama.cpp, which reads split GGUF files directly. For serving many users, the model card recommends SGLang or vLLM on the FP8 weights.

A single 24 GB or 48 GB consumer GPU cannot hold any of these. If that is your hardware, use the cloud tag or the smaller GLM 5.3 Flash.

FAQ

Is GLM 5.3 on Ollama? Yes, as glm-5.3:cloud. It runs on Ollama's servers, so it needs an Ollama account and API key.

Can I ollama pull glm-5.3 and run it offline? No. The Ollama library has no local tag for GLM 5.3. For offline use, download the weights or a GGUF quant from Hugging Face and run them with llama.cpp, SGLang or vLLM.

How much memory does GLM 5.3 need locally? The smallest GGUF quant is about 217 GB, the 4-bit quant about 467 GB, and the official FP8 weights about 753 GB.

What does GLM 5.3 cost on Ollama? Ollama's page lists $1.40 per 1M input tokens, $0.26 cached and $4.40 output, the same as Z.AI's API.

Can I use GLM 5.3 without any setup? Yes. You can chat with GLM 5.3 in the browser, or call Z.AI's API with the model ID glm-5.3.


Sources

Last updated: November 11, 2026

Start Using GLM 5 Today

Try GLM 5 free — reasoning, coding, agents, and image generation in one platform.