On this page
GLM 5.3 Flash is a 320B model, and its smallest build is a 93.1 GB file that needs about 100 GB of memory, so only a 128 GB Mac, a 128 GB workstation or a GPU server can run it. If you don’t have that, use it through an API that doesn’t keep your prompts, such as DeepInfra or Fireworks at $0.15 per million input tokens, or a flat-price plan. On a normal laptop, Off Grid AI Desktop runs smaller open models for free, fully offline.
Download Off Grid AI Desktop (free, Mac, Windows and Linux) or see the Boost plans →
What GLM 5.3 Flash is
GLM 5.3 Flash is Z.ai’s smaller GLM 5.3. Its model card lists 320B total parameters with 18B active per token, an MIT licence and image input. It has a context of up to 1,048,576 tokens and three thinking levels: low, high and max.
It’s the third most used model on OpenRouter this month. The 18B active parameters make it quick once it’s loaded, but all 320B still have to fit in memory.
What it takes to run it yourself
Unlike DeepSeek V4.1 Flash and Qwen3.8-Flash-Next, GLM 5.3 Flash has no lookup table you can leave on the SSD. The whole file has to fit in RAM plus VRAM, or unified memory.
| Unsloth build | File | 128 GB machine | 256 GB machine |
|---|---|---|---|
| 1-bit (UD-IQ1_S) | 93.1 GB | Tight | Runs well |
| 2-bit (UD-IQ2_XXS) | 101.8 GB | Tight | Runs well |
| 2-bit (UD-Q2_K_XL) | 108.7 GB | Too big | Runs well |
| 3-bit (UD-IQ3_XXS) | 120.4 GB | Too big | Runs well |
| 3-bit (UD-Q3_K_XL) | 147.5 GB | Too big | Runs well |
| 4-bit (UD-Q4_K_XL) | 199.7 GB | Too big | Tight |
File sizes are from Unsloth’s run guide, checked 16 November 2026. The ratings use the rule Off Grid AI uses on a computer: a file up to 65% of memory runs well, up to 82% is tight. Unsloth’s own advice is looser. It says 1-bit works on 100 GB and that it runs the 120 GB 3-bit build on 128 GB devices like a Mac or an NVIDIA DGX Spark.
In Unsloth’s tests, the 1-bit build agrees with the full model’s top answer 71% of the time, 3-bit (UD-IQ3_XXS) 82% and 4-bit 92%.
Which machines that means
- Laptops. The largest MacBook Pro has 128 GB, so it’s the only common laptop that can hold the 1-bit build. Everything at 64 GB or less is out.
- Desktops. A 128 GB Mac Studio or a PC with 128 GB of RAM runs the 1-bit build. For 3-bit or 4-bit, you need 256 GB. Our guide to what Mac you need has dated Apple prices.
- GPU servers. On one NVIDIA B200, Unsloth measured about 63 tokens a second writing with the 1-bit build.
Run it with llama.cpp
Unsloth’s guide builds a specific llama.cpp branch with GLM 5.3 Flash support. Use the build steps from the guide, with CUDA on for an NVIDIA machine and off for a Mac. Then download a build that fits:
pip install -U "huggingface_hub[cli]"
hf download unsloth/GLM-5.3-Flash-GGUF \
--local-dir unsloth/GLM-5.3-Flash-GGUF \
--include "*UD-IQ1_S*"
Start it as a server, using Unsloth’s sampling settings:
./llama.cpp/llama-server \
--model unsloth/GLM-5.3-Flash-GGUF/UD-IQ1_S/GLM-5.3-Flash-UD-IQ1_S-00001-of-00003.gguf \
--temp 1.0 \
--top-p 0.95 \
--chat-template-kwargs '{"reasoning_effort":"high","clear_thinking":true}' \
--host 0.0.0.0 --port 8080
clear_thinking drops old reasoning from the chat, which Z.ai recommends for normal conversation. Use "max" for hard problems; it’s the model’s default.
Chat with it from your laptop
Your laptop doesn’t need to hold the model. Off Grid AI Desktop can use the big machine’s llama-server as a remote model:
- Open Settings and find Remote model server.
- Under Saved servers, click Add server.
- Type a Server name and the Address, for example
http://192.168.1.30:8080. - Click Test connection, pick the model under Text + Vision, then click Save.
It shows up in the chat’s model picker with a Remote tag, and your prompts stay on your own network. See how to use a remote llama.cpp server from Off Grid AI Desktop for the details.
We haven’t tested GLM 5.3 Flash inside Off Grid AI Desktop’s own engine.
Use it through an API
Most people should use GLM 5.3 Flash from a host. Many hosts charge the same price, so pick on what they do with your prompts.
| Host | Price per 1M tokens (in / out) | Stores prompts? | Trains on them? |
|---|---|---|---|
| DeepInfra | $0.15 / $0.50 | No | No |
| Fireworks | $0.15 / $0.50 | No, zero retention by default | Not stated |
| Ollama Cloud | Ollama’s cloud plans | Zero data retention; US and Europe | No |
| Together | $0.15 / $0.50 | Yes, by default | Off by default |
| Z.ai API | $0.15 / $0.50 | Yes, while you have an account; Singapore | Lists training and improving its models among its uses |
| OpenRouter | From $0.04 / $0.08 | Zero-retention routing is a setting | No |
Host prices are as listed on OpenRouter, checked 16 November 2026. OpenRouter’s cheapest providers serve 4-bit (FP4) builds, so a lower price can mean a more compressed model.
Want a flat price instead of a meter?
Off Grid AI Boost is a flat-price API for open models on our own GPUs. The plan list names Qwen 3.8 and DeepSeek V4.1 Flash “and more”. GLM 5.3 Flash isn’t named, so check the Boost model list at launch. If DeepSeek V4.1 Flash does the job for you, it’s on the list today.
Boost starts at $4.99 a month for the first cohort, with regular prices from $9.99 to $199.99. We don’t store your prompts or train on them. That’s our promise today, and the full policy comes with launch. Usage limits for each plan aren’t published yet.
Boost starts in January; pre-order to lock in half price. API keys arrive by January 2027.
What runs on the computer you have
If you want an open model that runs offline today, use Off Grid AI Desktop. It’s free and needs no account.
| Your memory | What to run | File |
|---|---|---|
| 16 GB | Qwen 3.8 27B as a 2-bit build, or a 9B model | Under 10.4 GB |
| 32 GB | Qwen 3.8 27B, UD-Q4_K_M | 16.56 GB |
| 64 GB | Qwen 3.8 27B, Q8_0 | 29.05 GB |
- Open Models, then Text.
- Type the model name in Search models…. Results from Hugging Face show under More from Hugging Face.
- Pick a file. Each one shows its size and whether it’s Tight or Too big for this computer.
- Choose Download, then Use when it finishes.
For 16 GB, how to run Qwen 3.8 on a 16 GB laptop walks through the best file.
What goes wrong, and the fix
| Symptom | Check | Fix |
|---|---|---|
| llama.cpp says the model type is unknown | Your llama.cpp build | Build the branch named in Unsloth’s guide |
| It loads, then the machine swaps | File size against your memory | Drop to UD-IQ1_S, or close other apps |
| Long reasoning before every reply | Reasoning effort | Set "reasoning_effort":"low" or "high" |
| Old reasoning fills the context | clear_thinking |
Pass "clear_thinking":true |
| Test connection fails in Off Grid AI Desktop | Whether llama-server uses --host 0.0.0.0 |
Restart it with --host and allow it through the firewall |
FAQ
How much RAM does GLM 5.3 Flash need?
About 100 GB of RAM plus VRAM, or unified memory, for Unsloth’s 1-bit build, which is a 93.1 GB download. Unsloth runs its 3-bit build on 128 GB machines. The 4-bit build is 199.7 GB.
Can a laptop run GLM 5.3 Flash?
Only one with 128 GB of memory, such as a MacBook Pro with M5 Max and 128 GB, and even then the 1-bit build is tight. A 16 GB, 32 GB or 64 GB laptop can’t hold it.
Which GLM 5.3 Flash API is private?
DeepInfra says it doesn’t store or train on your data, Fireworks says zero retention by default, and Ollama Cloud says its GLM 5.3 Flash follows a zero data retention policy. Z.ai’s own API keeps your input while you have an account and may use data to improve its models.
Does Off Grid AI Boost serve GLM 5.3 Flash?
Boost names Qwen 3.8 and DeepSeek V4.1 Flash on its plan list, not GLM. Check the Boost model list at launch. Boost is a pre-order with API keys by January 2027.
Is GLM 5.3 Flash open source?
The weights are open under the MIT licence on Hugging Face, so you can run, change and use them commercially.
Related guides
- How to Run DeepSeek V4.1 Flash in 2026 (What It Really Takes, and the Private Option)
- What Mac Do You Need to Run Qwen 3.8 or DeepSeek in 2026?
- Best Private LLM APIs in 2026 (Prompts Never Stored or Trained On)
- Best Flat-Rate APIs for Open Models in 2026 (No Per-Token Bills)
Check what your own computer can run before you rent anything: download Off Grid AI Desktop.