Google Gemma 4 model demonstrates frontier-class AI inference capability running on a single GPU.

Google DeepMind launched Gemma 4 this week, releasing four open-weight models that fit entirely on a single 80GB Nvidia H100 GPU while delivering benchmark scores that rival models 20 times their size. The model family spans four sizes, with a 31-billion-parameter dense transformer currently ranking third among all open models on the Arena AI text leaderboard with an estimated score of 1452, and a 26-billion-parameter mixture-of-experts variant that activates only 3.8 billion parameters during inference and secured the sixth spot on the same leaderboard. Two smaller "effective" models at 4 billion and 2 billion parameters target smartphones, Raspberry Pi boards and Nvidia Jetson Orin Nano edge devices where battery life and memory constraints dominate design decisions. The release marks Google's most aggressive move yet against Meta's Llama in the open model race and hands Nvidia a new reason to sell GPUs to enterprises that want to run AI locally rather than pay per-token cloud fees.

Beyond parameter counts, Gemma 4 introduces capabilities that matter for enterprise AI workflows. All four models support native function calling and structured JSON output, enabling developers to build autonomous agents that interact with external tools without additional prompt engineering. Context windows extend to 256,000 tokens for the larger models and 128,000 tokens for the edge variants, making it possible to process entire codebases or large document sets in a single prompt. The models natively process images and video across all four sizes, while the two smaller edge models add native audio input for speech recognition and understanding directly on device. Google trained the family on more than 140 languages, positioning Gemma 4 as a practical option for multinational deployments where a single model needs to handle diverse language requirements.

The most consequential change has nothing to do with model architecture. Google shipped Gemma 4 under an Apache 2.0 license, abandoning the restrictive custom license that governed previous Gemma generations. That shift removes commercial use restrictions and acceptable-use policy enforcement that previously forced enterprise legal teams to review every deployment. Nvidia moved quickly to position Gemma 4 across its entire product line, from Blackwell data center GPUs down to Jetson edge modules and consumer GeForce RTX cards, with the 31-billion-parameter model running unquantized in BF16 precision on a single H100, while quantized versions fit comfortably on consumer GPUs with 24GB of memory. AMD also announced day-zero support for Gemma 4 across its Instinct data center GPUs, Radeon workstation GPUs and Ryzen AI processors. Enterprise deployment runs through Nvidia's NIM microservices offering prepackaged inference containers and the NeMo Automodel library for fine-tuning directly from Hugging Face checkpoints, supporting supervised fine-tuning and LoRA techniques without requiring model conversion.

Gemma 4 enters the most competitive open model market in the industry's history. Meta's Llama 4 Scout offers a 10-million-token context window, while Alibaba released Qwen 3.6-Plus on the same day with a one-million-token context window. Chinese competitors including DeepSeek, Moonshot AI and Z.AI continue to release models that rival proprietary frontier systems. The generational improvement over Gemma 3 is substantial—on the AIME 2026 math competition benchmark, the 31-billion-parameter model scores 89.2% compared to 20.8% for the previous 27-billion-parameter Gemma 3. Google's strategic advantage lies in the combination of strong benchmarks, permissive licensing that matches Qwen's openness and exceeds Llama's more restrictive community license, and broad hardware support across the most widely deployed GPU ecosystem in enterprise data centers. Early community testing within 24 hours of release, however, surfaced concerns about inference speed, particularly with the 26-billion-parameter mixture-of-experts model, with developers reporting that real-world throughput fell short of expectations on some hardware configurations.