Knowledge Base · Local AI
How much VRAM to run Qwen 3 locally?
By Tiny · how I pick
As an Amazon Associate I earn from qualifying purchases. The graphics card links below are affiliate links.
At 4-bit, the Qwen 3 models need about:
- 8B: 6 GB
- 14B: 10 GB
- 32B: 20 GB
The interesting one is Qwen 3 30B-A3B, a Mixture-of-Experts model. The whole model takes about 18 GB at 4-bit, but only about 3B of it works on each word, so it stays fast even when part of it sits in system RAM. That lets it run on an 8 to 12 GB card with the overflow in RAM, and it reasons like a much bigger model.
Words you'll see in this guide
- VRAM: the memory on your graphics card. AI models load into it, so more VRAM means you can run bigger models.
- Mixture-of-Experts (MoE): a model design that only uses part of itself for each word it writes, so it runs faster than its total size suggests. The whole model still has to be loaded.
- Offloading: running part of a model on the graphics card and the rest on the processor and system RAM. It works, but slowly.
- Context: how much text the model can keep in mind at once, measured in tokens. Longer context uses more memory.
VRAM by Qwen 3 model
| Model | Memory (4-bit) | Runs on |
|---|---|---|
| Qwen 3 4B | About 3 GB | Almost any graphics card, or the processor alone |
| Qwen 3 8B | About 6 GB | An 8 GB card |
| Qwen 3 14B (best balance) | About 10 GB | A 16 GB card, with long conversations |
| Qwen 3 30B-A3B (MoE) | About 18 GB total | Fully on a 24 GB card, or an 8 to 12 GB card with the rest in system RAM |
| Qwen 3 32B | About 20 GB | A 24 GB card like the RTX 3090 or RTX 4090 |
- 16 GB card: Qwen 3 14B for the best all-round experience, or the 30B-A3B with a little of it in system RAM for stronger reasoning.
- 8 GB card: the 8B. The 30B-A3B also works if you have 32 GB of system RAM for the overflow.
- 24 GB card: the 32B, which is especially strong at code.
How to run it
Use Ollama with Open WebUI. Not sure what fits? The Server Calculator gives you a straight answer. Then:
- Install Ollama (the guide above walks through it).
- Download and start Qwen 3:
ollama run qwen3You should see: a
>>>prompt once the download finishes. - Type a question. Add Open WebUI if you want a ChatGPT-style window in your browser.
Where to go next
- Need a card? See the best GPU for local AI under $500 and what a 16 GB vs 24 GB card runs.
- Comparing model families? See VRAM for DeepSeek, VRAM for Llama 4 and VRAM for Gemma 4.
Quick answers
How much VRAM do I need to run Qwen 3 locally?
At 4-bit, Qwen 3 8B needs about 6GB, the 14B about 10GB, and the 32B about 20GB. Qwen 3 30B-A3B needs about 18GB in total, and because it's fast, it also runs well on smaller cards with part of it in system RAM.
What is Qwen 3 30B-A3B and why is it so efficient?
It's a mixture-of-experts model with 30B parameters in total but only about 3B working at a time. It answers like a bigger model while running at the speed of a small one. It still needs memory for all 30B, about 18GB at 4-bit.
Which Qwen 3 model should I run on a 16GB GPU?
Qwen 3 14B fits comfortably with room for long chats. The 30B-A3B also works on 16GB with a little of it in system RAM. On a 24GB card, step up to Qwen 3 32B.
Can Qwen 3 run without a dedicated GPU?
Yes. The 4B and 8B run on a processor with enough system RAM, just slower, and the 30B-A3B is usable on strong processors and Apple Silicon Macs because so little of it works at once.
Free email updates
Qwen ships new versions often. Subscribe and I'll tell you when a new one fits your card better.
No spam. Unsubscribe anytime. Privacy policy.