Gemma 4 E2B It

Overview. Gemma 4 E2B is an instruction-tuned, open-weights model from Google DeepMind, released under the Apache 2.0 licence. It has 5.1B total parameters, with 2.3B effective parameters doing the compute, and accepts text and image input with a 128K token context window. Developers reach for it when they need a compact multimodal model that can also reason before it answers.

Key strengths. No benchmark scores for this variant are in the catalogue yet, so this section makes no quality claims. The model's strengths are its structure: a compact dense footprint of 2.3B effective parameters, a 128K token context window, and built-in reasoning that the caller can switch on or off with a thinking toggle.

When to use it. Use this model when you need image understanding alongside text in a small, dense model, or when a long document or multi-turn session needs to fit in a single request of up to 128K tokens. Turn the thinking toggle on for multi-step questions where working through the problem matters more than a direct reply. Leave it off for simple, direct tasks.

Limitations. The model outputs text only, and this record does not yet include benchmark results, so you should test it against your own workload before relying on it. Its pre-training data runs to January 2025. The creator lists audio input, and the Amazon Bedrock model card also lists audio as an input modality, on the Responses API.

Sovereignty note. The weights are Apache 2.0 licensed. Through Stav it is served by Amazon Bedrock in the AWS European Sovereign Cloud, in the eusc-de-east-1 region in Germany, with no cross-region inference. Data retention is set to none on Stav's account, and AWS's European Sovereign Cloud addendum bars it from using customer data beyond providing the service. The operator is a subsidiary of a US company, so US lawful-access law can reach it.