Gemma 4: Google launches Gemma 4 on Good Friday, delivering standout performance metrics.

Google said Gemma 4 tops the Arena AI leaderboard, ranking #3 among dense models with just 31 billion parameters, according to the Pudgy Cat report. The same analysis places the 26 billion‑parameter mixture‑of‑experts (MoE) version at #6 globally, a performance the outlet describes as “unprecedented intelligence per parameter.”

The launch bundles four variants under an Apache 2.0 license, a shift highlighted by AINews’ community roundup. The smallest, Gemma 4 E2B (2 B parameters), runs on smartphones, Raspberry Pi and Jetson Nano with “near‑zero latency” and native audio support, while the 4 B‑parameter E4B model targets the same edge devices with higher capability. Both are marketed as “Effective” models engineered to punch above their weight on constrained hardware, per the Pudgy Cat summary.

Google’s internal research stack, the same that powers Gemini 3, underpins the new family, the report notes. The 31 B dense model allegedly outperforms rivals “20 times larger,” a claim echoed by the same source as the core selling point of the release.

Day‑0 ecosystem support arrived across major runtimes, AINews reported. vLLM added GPU, TPU and XPU back‑ends simultaneously; llama.cpp, Ollama and Intel’s Xeon, Xe GPU and Core Ultra stacks all listed immediate compatibility. Unsloth announced local run and fine‑tune tools, and Hugging Face’s Infer platform posted ready‑to‑deploy endpoints.

Industry voices on social media praised the move. François Chollet called Gemma 4 “Google’s strongest open model yet” and urged developers to use the JAX backend in KerasHub, while Demis Hassabis highlighted its efficiency, saying it “outperforms models 10× larger on Google’s chart.” The license shift was called a “real open‑weights release” by Clement Delangue, QuixiAI and the official @googlegemma account, according to the AINews roundup.