Gemma 4 E2B
Model, local llm, by Google
Assessment
Tried as the small helper and set aside. It reasons by default, which the OpenAI-style endpoint cannot switch off, its file does not fit a 6 GB card, and its vision misread a plain photo. The Qwen 3B pair does the same jobs with fewer surprises.
2026-10-09
Strengths
- A capable small chat model with a long context
- Runs on a phone-class device in its intended setting
Limitations
- A thinking model: hidden reasoning unless turned off through Ollama's native API
- Its file is larger than a 6 GB card, so it spills
- Vision proved unreliable on an older NVIDIA card
Gemma 4 E2B is the smallest Gemma 4 model, built for on-device use: about two billion effective parameters in an architecture that shares embeddings to keep memory low.[1] It is a thinking model, meaning it writes a hidden chain of reasoning before each answer unless told not to, and it has a vision input.
Why it was tried
As the text helper on a desktop card: small, modern, multimodal. On paper it covers both the text and the photo jobs with one model.
Why it was set aside
Three things, each a practical problem rather than a quality one.[2] First, Ollama's OpenAI-compatible endpoint drops the chat_template_kwargs field, so the only way to stop the hidden reasoning is Ollama's native API with think: false; every client has to know that. Second, its file is larger than the card it was meant for, so part of it spills to system RAM and speed suffers. Third, on a GTX 1060 its vision answer to a photo with a red square beside a printed total was "overlapping ID cards", which is not a reader to trust.
Where it would fit
A newer card with more memory, or a phone, where its design pays off. On a 6 GB desktop card the Qwen2.5 3B pair, one for text and one for photos, is the steadier choice.
History
- 2026-10-09: article written and published.
- 2026-10-09: entry created.
Comments
Public comments on each entry are coming. Nothing is collected here yet.