Gemma 4 12B Unified: Open Weights Go Multimodal
What was announced?
On June 3, 2026, Google released Gemma 4 12B Unified, the latest entry in its open-weight model family. Three details stand out on the release page. The weights ship under the Apache 2.0 license, which allows commercial use and modification. The model is unified across modalities, accepting text, audio, and image input in one set of weights. And it carries a 256K-token context window, enough to hold a long document collection in a single pass.
The size is the strategic part. Twelve billion parameters, quantized, runs on a well-equipped laptop. This is not a phone model, and it is not a data-center model. It is sized for hardware many people already own.
What does it mean for personal, on-device writing?
“Private AI” has largely been Apple’s story: small models, deeply integrated, running on the device in your hand. Open releases like this widen the story beyond Apple’s walls. Anyone with a capable laptop can now run a multimodal assistant locally, with no account and no server involved.
Precision matters here. A laptop is not a phone. A 12B multimodal model wants more memory than phone-class silicon offers, which is why Apple Intelligence relies on far smaller foundation models tuned for iPhone hardware. Gemma 4 12B Unified does not change what runs in your pocket.
The direction still matters more than the size class. Every credible open release pushes the same norm: capable AI does not need to phone home. First it becomes normal on laptops. Then people start expecting it everywhere, including the phone. Norms travel down the hardware ladder faster than models do.
Where does ILURA stand?
This release changes nothing inside the app. ILURA does not load third-party weights. Drafting runs on Apple’s on-device foundation models through Apple Intelligence, the same as yesterday.
But ILURA still benefits. The app’s premise is that personal writing AI should live where the writing happens: on the iPhone, with no account to create, under a privacy label that reads Data Not Collected. That premise gets stronger every time a serious local model ships, from any vendor. Google publishing laptop-runnable multimodal weights under Apache 2.0 tells the market that local is a real deployment target, not a hobbyist niche.
Your role playbooks, your corrections, and your drafts stay on the device either way. The industry is just making that arrangement look less unusual by the month.
Read the signal through ILURA
Platform news matters when it changes what users expect from personal AI. ILURA reads these shifts through one lens: private agents trained by the user on iPhone.
- What becomes possible?
- What should stay user-controlled?
- What belongs on device?
Try it now
Put this to work on a real message.
Open ILURA, bring in a message you actually need to handle today, and get it done in your voice — free, on device, no account. It learns the preference, so the behavior carries to the next one.
Free to start · No account · Data Not CollectedQuick answers
- Can Gemma 4 12B Unified run on a phone?
- No. A 12-billion-parameter multimodal model targets laptops and desktops with enough memory. Phone-class on-device AI still depends on much smaller models, like the ones behind Apple Intelligence.
- Does ILURA use Gemma models?
- No. ILURA drafts with Apple's on-device foundation models through Apple Intelligence. Open releases like Gemma still matter, because they normalize local, private AI across the industry.