Gemma 4: Google’s New Open Models Optimized for NVIDIA GPUs

Google and NVIDIA have deepened their collaboration to optimize Google’s new Gemma 4 family of open models for NVIDIA GPUs, enabling efficient performance across a wide range of systems – from data centers to personal AI supercomputers and edge AI modules. The move aims to extend AI innovation beyond the cloud to everyday devices, leveraging local, real-time context for more effective AI applications.

Announced on April 2, 2026, Gemma 4 introduces a class of small, fast, and versatile models designed for efficient local execution. Google reports that developers have downloaded Gemma models over 400 million times, building a community of more than 100,000 variants. Gemma 4 is released under an Apache 2.0 license.

Gemma 4 Model Variants and Capabilities

The Gemma 4 family includes four variants: Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE), and 31B Dense. These models are designed for a range of tasks, including reasoning, coding, agentic workflows, and multimodal interactions. According to Google, the 31B model currently ranks as the #3 open model globally on the Arena AI text leaderboard, while the 26B model holds the #6 spot.

• Reasoning: Strong performance on complex problem-solving tasks.
• Coding: Code generation and debugging for developer workflows.
• Agents: Native support for structured tool use (function calling).
• Vision, Video and Audio Capabilities: Enables rich multimodal interactions for object recognition, automated speech recognition, and document or video intelligence.
• Interleaved Multimodal Input: Mix text and images in any order within a single prompt.
• Multilingual: Out-of-the-box support for 35+ languages, pretrained on 140+ languages.

The E2B and E4B models are optimized for ultra-efficient, low-latency inference at the edge, capable of running offline on devices like NVIDIA Jetson Orin Nano modules. The 26B and 31B models are designed for high-performance reasoning and agentic AI, running efficiently on NVIDIA RTX GPUs and the NVIDIA DGX Spark.

NVIDIA Support and Local Deployment

NVIDIA has collaborated with Ollama, and llama.cpp to provide a streamlined local deployment experience for Gemma 4. Users can download Ollama to run the models or install llama.cpp and pair it with the Gemma 4 GGUF Hugging Face checkpoint. Unsloth also provides optimized and quantized models for efficient local fine-tuning and deployment via Unsloth Studio.

NVIDIA emphasizes that running open models like Gemma 4 on NVIDIA GPUs achieves optimal performance due to the acceleration provided by NVIDIA Tensor Cores for AI inference workloads. The CUDA software stack ensures broad compatibility across frameworks and tools, facilitating efficient model execution.

This collaboration allows Gemma 4 to scale across a wide range of NVIDIA systems, from Jetson Orin Nano at the edge to RTX PCs, workstations, and DGX Spark, without requiring extensive optimization.

Agentic AI and OpenClaw Integration

As local agentic AI gains momentum, applications like OpenClaw are enabling always-on AI assistants on RTX PCs, workstations, and DGX Spark. The latest Gemma 4 models are compatible with OpenClaw, allowing users to build capable local agents that leverage personal files, applications, and workflows to automate tasks. NVIDIA has also introduced NVIDIA NemoClaw, an open-source stack that optimizes OpenClaw experiences on NVIDIA devices by increasing security and supporting local models.

Accomplish.ai has also announced Accomplish FREE, a no-cost version of its open-source desktop AI agent with built-in models, harnessing NVIDIA GPUs for fast, private, and zero-configuration execution.

Users interested in getting started can find more details on the NVIDIA technical blog and the Google DeepMind announcement blog.