Gemma 4 E4B: the biggest, and why we left its sampling alone
5.2GB of weights and 2GB of GPU buffer for the best quality we can run in a tab. Plus an afternoon lost to lowering a temperature we should not have touched.
Gemma 4 E4B is the biggest model of our catalogue and, let us say it immediately, it is not for everybody. 5.2GB of weights to download, 7GB of working set, a minimum GPU buffer of 2GB. If you are reading this on a laptop, there is a good chance it will not load.
For those whose machine follows, it is the best quality we can offer in a browser tab today.
The numbers
| Budget | Value |
|---|---|
| Download | ~5.2GB |
| Working set | ~7GB |
| Minimum GPU buffer | 2GB |
| Precision | q4f16 |
| Runtime | gemma4 |
The 2GB buffer is the hard limit. It is not a recommendation, it is a condition: below, the allocation fails and the model does not load at all. A GPU with 8GB will pass, a GPU with 4GB will not, whatever the rest of the configuration. The AI Readiness Test verifies it before anything is downloaded.
A different sampling
Google publishes a preset which is very far from that of Qwen:
{ "temperature": 1.0, "top_p": 0.95, "top_k": 64 }
Temperature at 1.0, top_k at 64. It is much hotter than everything else in
our catalogue, and one's first reflex is to lower it. We did it. It was a
mistake: the model becomes flat and repetitive, and the quality of the code
drops. Gemma has been post-trained with this sampling, the distribution is
calibrated for it. So we leave it.
The lesson, which cost us an afternoon: a decoding preset is not a global setting that one tunes to taste. It belongs to the model, it is published with it, and one changes it only with a measurement to justify it.
Its own runtime
Like the bigger Qwen, Gemma 4 passes through a dedicated runtime, gemma4,
with its own KV cache management. The reason is the same: without this, each
turn of a long conversation re-prefills everything, and on a model of this size
the wait becomes unbearable.
Who is it for
If you have a recent desktop GPU, this is the model which gives the best answers of the site, on long reasoning, on code that involves several files, on anything which asks to hold a lot of context at once.
If you are not sure that your machine follows, do not start with this one. Qwen3.5 2B is at 1.6GB and covers most of the needs, and Qwen3.5 4B is a good intermediate at 3GB.
And as always: 5.2GB pass through your GPU, nothing passes through our servers. The Local LLM workspace sends no prompt anywhere.