Google's Gemma 4 family is the company's biggest open-weight commitment to date. Four sizes ship under Apache 2.0: a 2B for on-device, a 9B for laptops, a 31B dense for desktop GPUs, and a 70B MoE for servers. The 31B Dense is the one that matters most for brand visibility, because it is competitive with closed models 20x its parameter count on instruction-following and tool-use benchmarks while running on a single consumer GPU.
What the Gemma 4 family covers
Gemma 4-2B for on-device deployment (mobile apps, browser extensions, IoT). Gemma 4-9B for laptops and small servers, optimized for sub-second latency. Gemma 4-31B Dense as the developer flagship: 80GB GPU friendly, MMLU 84.7%, IFEval 87.2%, SWE-Bench Verified 64.3%, beats Mistral Large 2 on most tasks at one-fourth the inference cost. Gemma 4-70B MoE for cloud serving with frontier-comparable performance at open-weight pricing.
Why this matters for brand visibility
Two distinct shifts. First, Gemma 4 is the model Google will push into Pixel devices, Android, ChromeOS, and the broader Google for Developers ecosystem. That means hundreds of millions of devices will run a Google-trained model directly, with brand recall driven by Google's training corpus (which heavily overweights Google Search results, Wikipedia, and YouTube transcripts). Brands strong in Google Search ranking translate well; brands strong in social-only or paid-only channels translate poorly.
Second, Gemma 4-31B Dense becomes the default for indie developers, hackathons, and bootstrapped AI startups. The reason is the price-performance curve: it runs locally without cloud cost, ships under Apache 2.0, and has tooling parity with the rest of the Gemma ecosystem. Every product built on Gemma 4 inherits Google's training cutoff and corpus. The compounding effect over the next 12 months is that "Google ecosystem brand recall" extends from Search and YouTube into thousands of derivative products built on Gemma.
What to test this week
Pull Gemma 4-31B Dense from Hugging Face or via Ollama and run brand-recall tests against your top three competitors. Compare against Gemini 3.1 Pro on the same prompts. If Gemma's answers diverge significantly from Gemini's, that is a fingerprint of how filtering and training-data selection differ between the open-weight and closed-weight Google stacks.