⚡️ Google's Open AI Strategy — Omar Sanseviero, Google DeepMind

Google’s Gemma 4 introduces an architectural tweak—per-layer embeddings used as lookup tables—that shifts many parameters off GPU into CPU/disk. Result: a 2B “active” model that effectively uses ~5B params without the VRAM bill. It’s optimized for phones and constrained devices, not for supersizing. Translation for teams: you can credibly ship privacy-preserving, low-latency features locally, but don’t expect state-of-the-art world knowledge.

Android Studio now supports an “agent mode” that can run Gemma 4 (or any OpenAI-compatible endpoint, local or remote). This is a real path for IP-sensitive orgs to get coding assistance without sending source to the cloud. Expect IDE-native, offline assistants to move from novelty to default for regulated and high-trust environments.

Multimodality is practical: images, audio, and short video understanding, but no image segmentation and no combined audio+video yet. The multilingual tokenizer (shared DNA with Gemini) is a quiet win: regional fine-tunes work with less data. However, partners report the base model often performs well enough to skip custom fine-tunes—prompting and retrieval are beating instruction tuning for many general tasks.

On architecture choices, Gemma ships both 31B dense and 27B MoE (≈4B active). Dense is easier to fine-tune and control; MoE is inference-efficient but finicky to fine-tune (routing, hyperparams, stability). If you need predictable post-training behavior, prefer dense. If you need throughput at fixed VRAM, MoE may help—treat as specialized infra, not a customization target.

Google is exploring diffusion transformers for text/code. The pitch is speed, but quality lags autoregressive and fine-tuning is harder. Consider it R&D; not a near-term replacement for production codegen, except possibly as a narrow “executor” inside an agent.

Operationally, beware on-device LoRA sprawl. Bundling multiple app-specific adapters explodes update complexity and drains battery. Standardize on a single base model per device profile with centralized distribution and lifecycle. Kaggle’s move into agent evaluations is noteworthy—expect more public leaderboards, but build task-grounded internal evals to avoid being gamed.