- Learn
- Pick a model
Qwen3 vs Gemma 3: which local model should you run?
Qwen3 and Gemma 3 compared for local use. Reasoning, images, writing, licenses, memory and estimated speeds on 12 and 24 GB cards.
Pick Qwen3 for reasoning, code and speed. Pick Gemma 3 if you need it to read images, or you write a lot in languages other than English. Both are good, both run well on ordinary hardware, and many people keep one of each.
The short version
| Qwen3 | Gemma 3 | |
|---|---|---|
| Made by | Alibaba (Qwen) | Google DeepMind |
| Sizes we track | 4B, 8B, 14B, 32B, 30B-A3B (MoE) | 12B, 27B |
| Reads images | No | Yes |
| Thinking mode | Yes, on or off per message | No |
| Context | 32K, 128K with YaRN | 128K |
| License | Apache 2.0 | Gemma Terms of Use |
| Best at | Reasoning, math, code, tool calling | Images, writing, translation |
Where Qwen3 wins
Reasoning and code. Qwen3 can think step by step before it answers. Add /think to a message for hard problems and /no_think when you just want a quick reply. On tricky code and math that makes a real difference, and Gemma 3 has no equivalent switch.
Speed for its smarts. Qwen3 30B-A3B is a mixture-of-experts model. It holds 30B parameters but reads only about 3B per token, so on a 24 GB card it runs several times faster than any dense model of similar quality. Gemma 3 has nothing like it.
More sizes. Qwen3 comes in a size for almost every card, from a 4B that runs on a laptop to a 32B that fills a 24 GB card.
A simpler license. Apache 2.0 is a standard open-source license. Gemma has its own terms. If that matters for your project, see open model licenses explained.
Where Gemma 3 wins
It reads images. Drop in a screenshot, a chart, a photo of a receipt or a page of handwriting and ask about it. Gemma 3 12B is the smallest model on our list that can do this. If you want vision, that settles it.
Writing and languages. Gemma 3 tends to write more naturally, and it is strong across many languages. For drafting, rewriting and translation, try it before you assume Qwen3 is better.
Long documents. It has a 128K context out of the box. Qwen3 needs YaRN switched on to go past 32K. Either way, long context costs memory; see how much VRAM you need.
Speed and memory, head to head
The pairs that compete on the same card, from our calculator:
| Card | Qwen3 | Gemma 3 |
|---|---|---|
| RTX 3060 12 GB | 14B: ~21 tok/s | 12B: ~24 tok/s |
| RTX 3090 24 GB | 14B: ~55 tok/s | 12B: ~64 tok/s |
| RTX 3090 24 GB | 32B: ~26 tok/s | 27B: ~30 tok/s |
| RTX 3090 24 GB | 30B-A3B: ~164 tok/s | none |
These are estimates. Size for size they are close, because both are dense models and speed follows file size. The outlier is Qwen3 30B-A3B, which is in a different league for speed.
Memory: Gemma 3 12B is an 8 GB download and Qwen3 14B is about 9 GB, so both fit a 12 GB card. Gemma 3 27B (17 GB) and Qwen3 32B (20 GB) both need a 24 GB card.
Our picks by card
- 8 GB: Qwen3 8B. Gemma 3 12B is too tight. See best local LLM for 8 GB.
- 12 GB: Qwen3 14B for thinking and code, Gemma 3 12B for images and writing. Both fit, so keep both.
- 24 GB: Qwen3 30B-A3B as the daily driver, Gemma 3 27B for images, Qwen3 32B for the hard ones. See best local LLM for 24 GB.
Try both in five minutes
Ollama only loads the model you're using, so having both on disk costs nothing but space.
ollama pull qwen3:14b && ollama pull gemma3:12bThen give each the same three prompts from your real work and compare. That beats any leaderboard. Our new model checklist has a quick way to do it.
Not sure what fits your machine? Start at every GPU.
One email when there's a new guide, podcast episode, or result from the rack. No spam.