Google Deploys Gemma 4 to Counter Chinese Open‑Weight Models, Boosts Vids with Veo and

Google’s latest Gemma 4 family arrives with a markedly broader feature set than its predecessors, positioning the model as a direct counter‑measure to the wave of open‑weight Chinese LLMs that have begun to rival GPT‑5 in benchmark tests. Developed by DeepMind, the fourth‑generation Gemma models ship under an Apache 2.0 license and support more than 140 languages, native function calling, and multimodal inputs for video and audio, according to the Register. The flagship 31‑billion‑parameter variant is tuned for “advanced reasoning,” a term Google uses to describe improved performance on chain‑of‑thought math problems and instruction‑following tasks. Unlike the larger, proprietary Gemini models, Gemma 4 is deliberately sized for on‑premise deployment: it can run unquantized at 16‑bit on a single 80 GB H100 GPU, while a 4‑bit quantized version fits comfortably on a consumer‑grade RTX 4090 or AMD RX 7900 XTX, making it accessible to enterprises that lack massive GPU clusters.

The rollout also integrates Gemma 4 into Google’s Vids platform, where the company has refreshed its video generation pipeline with the Veo 3.1 model and the newly announced Lyria music‑creation suite. Ars Technica notes that Veo 3.1, first deployed in Gemini late last year, delivers eight‑second, 720p clips with higher realism and temporal consistency than earlier iterations. Vids now offers three tiers of AI access: a free tier limited to ten video generations per month, an AI Pro subscription granting fifty clips, and an “AI Ultra” plan—available to both personal and enterprise customers—allowing up to 1,000 videos monthly. The Lyria models, meanwhile, can produce background scores without requiring lyric input, expanding Vids’ creative toolbox beyond visual content.

From a deployment perspective, Gemma 4’s hardware efficiency is a central selling point. The Register reports that the 31‑billion‑parameter model can be run unquantized on a single 80 GB H100, a configuration that would cost enterprises upwards of $30,000 in cloud GPU time for continuous inference. At 4‑bit precision, the same model shrinks to roughly 8 GB of memory, enabling it to run on a 24 GB consumer GPU with frameworks such as TensorRT or the open‑source DeepSpeed‑ZeRO. This scalability means that small‑to‑medium businesses can host the model on‑premise, preserving proprietary data and avoiding the data‑leak concerns that have plagued other open‑weight offerings. Google emphasizes that the model does not harvest user data for future training, a claim designed to differentiate Gemma 4 from Chinese competitors that have been criticized for re‑training on client inputs.

Function calling and multimodal capabilities further distinguish Gemma 4 from its rivals. Native function calling allows the model to invoke external APIs directly from generated code, streamlining workflows for developers building agentic AI assistants or automated coding tools. Video and audio inputs enable the model to process spoken commands, interpret visual scenes, and generate context‑aware responses—a feature set that aligns with Google’s broader strategy of integrating LLMs into its ecosystem of products, from Search to Workspace. The Register notes that the model’s multilingual breadth—covering over 140 languages—targets enterprise customers operating in diverse regions, offering a domestic alternative to Chinese models that have historically been limited to Mandarin and a handful of other languages.

The strategic timing of Gemma 4’s release reflects Google’s assessment of the competitive landscape. As Chinese firms such as Moonshot AI, Alibaba, and Z.AI push open‑weight LLMs that claim parity with GPT‑5 and Anthropic’s Claude, Google is leveraging its deep‑learning infrastructure and licensing flexibility to retain enterprise market share. By bundling Gemma 4 with Vids, Veo 3.1, and Lyria, Google creates a cohesive AI creation stack that can be adopted without the steep capital expenditures required for larger proprietary models. The combined offering promises a “domestic” solution that keeps sensitive corporate data in‑house while still delivering state‑of‑the‑art multimodal performance—a proposition that, according to the Register, could be decisive for organizations wary of data sovereignty issues.