Google DeepMind Launches Gemma 4 Amid Competition from Chinese Open Models

Google DeepMind, on April 2, launched Gemma 4, a new family of open models for reasoning tasks and agentic workflows. It follows earlier versions of Gemma, which have been downloaded more than 400 million times by developers.

Released under an Apache 2.0 license, the company said Gemma 4 delivers higher performance relative to their size and runs efficiently across devices ranging from mobile hardware to GPUs.

“Gemma 4 delivers an unprecedented level of intelligence-per-parameter,” said Google in a blog post.

The release includes four model variants — Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE) and 31B Dense.

The models are available through platforms including Hugging Face, Kaggle, and Ollama, with integration support across tools such as Transformers, vLLM, and llama.cpp. Developers can also deploy workloads on Google Cloud infrastructure.

According to the company, the 31B Dense model ranks third among open models on Arena AI’s leaderboard, while the 26B version ranks sixth, competing with significantly larger models.

The 31B dense model records an Elo score of around 1452, placing it close to Qwen3.5-397B-A17B at roughly 1450, and just a few points behind GLM-5 at about 1457 and Kimi-K2.5-Thinking at approximately 1456.

“Google just re-entered the game. They want to take the crown back from Chinese open source AI,” said Clement Delangue, CEO of Hugging Face, in a post on X.

The release comes as Chinese open models such as Qwen 3.5, GLM-5, and Kimi K2.5 gain traction globally,  with a focus on reasoning, coding and agentic capabilities.

Google said its E2B and E4B models run on-device with multimodal capabilities and low-latency processing. Larger models run on GPUs, including configurations using a single NVIDIA H100, with quantised versions available for consumer hardware. 

The models can be fine-tuned efficiently across Android devices, laptop GPUs, developer workstations and accelerators for research and production use.

The company said it worked with hardware partners such as Qualcomm Technologies and MediaTek to run it on mobile phones and IoT devices.

Google just re-entered the game 🔥🔥

They want to take the crown 👑 back from Chinese open source AI.

And... Gemma 4 is FINALLY Apache 2.0 aka real-open-source-licensed.

From what I've seen it's going to be a pretty significant model. But give it a try yourself today:

brew… pic.twitter.com/xQsoQDLxYw

— clem 🤗 (@ClementDelangue) April 3, 2026

Google said Gemma 4 shows improvements in multi-step reasoning, including math and instruction-following benchmarks. 

The models support agentic workflows with native function calling, structured JSON output and system instructions, enabling interaction with tools and APIs. They also generate code for offline use, allowing local development workflows.

The models process images and video across resolutions, with capabilities for tasks like OCR and chart understanding, while smaller variants also support audio input for speech recognition. 

Moreover, Gemma 4 handles long-form inputs with context windows of up to 256K tokens in larger models and 128K in edge versions. The models support more than 140 languages and can be fine-tuned for domain-specific tasks.

The company said the Apache 2.0 license allows developers to build and deploy across environments, including on-premises and in the cloud.