Gemma
last releaseApril 2, 2026
powered byGemma 4 31B, Gemma 4 26B MoE, Gemma 4 E4B
goblin vibe check:
good call if you're already in the google ecosystem and want something you can actually run locally
google's open model family for local, edge, mobile, and developer-controlled deployment. current gemma rows matter for indie builders because they cover multimodal use, function calling, structured output, and smaller edge variants without forcing everything through a hosted gemini product.
context
256k
tokens
Apache 2.0 fully open releaseMultimodal across text, image, and videoNative function calling and structured JSON outputEdge variants support native audio input
key features
Apache 2.0 fully open releaseMultimodal across text, image, and videoNative function calling and structured JSON outputEdge variants support native audio inputApache 2.0 open releaseedge variants target phones, Raspberry Pi, and low-latency devices
spec & usage
31B Dense ranked #3 on Arena AI at release while the 26B MoE runs with only 3.8B active parameters per inference step
E2B and E4B variants target phones, Raspberry Pi, and near-zero-latency edge workloads
Weights are available through Hugging Face, Kaggle, Ollama, and Google AI Studio
31B dense model and 26B MoE model target stronger desktop/cloud workloads
E2B and E4B variants target lightweight edge deployment
limitations
smaller edge models trade quality for deployment footprint
google's hosted Gemini models remain stronger for many frontier workloads
scope:
languagevisualmodelcodesearchagentapilocalopen-sourcebenchmark-strong
launchApril 2, 2026
last releaseApril 2, 2026