Google DeepMind Releases Gemma 4 12B, Beijing Plans $295 Billion AI Investment, Apple Unveils CoreAI Inference Engine

Google DeepMind Releases Gemma 4 12B

On June 9, Google DeepMind released a new model in the Gemma 4 family, Gemma 4 12B. A mid-size model that fills the gap between the edge-oriented E4B and the MoE-based 26B, its standout feature is a **unified architecture that requires no encoders**. By feeding image and audio inputs directly into the LLM itself instead of routing them through a multimodal encoder, it slashes the memory footprint while maintaining strong performance. It is also the first model in the Gemma 4 series to support **native audio input**.

Downloads across the Gemma 4 series have already surpassed 150 million. QAT (Quantization-Aware Training) optimized versions are also available on Hugging Face via unsloth; the quantized variants dramatically reduce memory requirements while keeping quality on par with bfloat16. Multiple variants are offered, including GGUF format and a mobile-optimized build (wNa8o8).

Unsurprisingly, the community quickly shipped an "OBLITERATED" edition as well. Built with a two-pass surgical pipeline that promises zero refusals and zero performance degradation, it is intended for alignment research and red-teaming use cases.

Beijing Plans a $295 Billion AI Infrastructure Buildout

Details have surfaced of China's massive AI infrastructure construction plan, backed by roughly $295 billion (about 43 trillion yen) in investment. The most notable point is a policy to **mandate that 80% of chips used be domestically produced**, effectively locking out US semiconductor suppliers. It is another step toward a sharper bifurcation of semiconductor supply chains in the US-China AI race.

Creative GPU sourcing is already underway in China — half-height PCIe cards retrofitted from used V100s have been circulating, for example — but if this plan goes ahead, raising the level of China's domestic semiconductor industry becomes an urgent task.

Apple Announces the CoreAI Inference Engine

At WWDC 2026, Apple announced **CoreAI**, the successor to CoreML. It is Apple's own on-device inference engine meant to replace MLX and llama.cpp, targeting optimization on iPhone and iPad in particular. For now, supported models are limited to those from mid-2025 and earlier, but a major expansion of ANE (Apple Neural Engine) operations is suggested.

Separately, Apple explained that privacy is preserved even when its AI runs on Google's servers, and an EU statement was reported indicating that the decision to hold back the Siri AI rollout in the EU was Apple's own. On top of that, a $250 million settlement has been proposed over allegedly exaggerated claims about iPhone 16 AI features.

Lovable Reaches $500 Million ARR

Lovable, an AI coding tool, announced that it has hit $500 million in annual recurring revenue (ARR). One million new projects are created every week, with users putting the tool to work building businesses and replacing internal software. The numbers back up just how fast the market for AI development tools is expanding.

AgentSploit: A Security Testing Tool for AI Agents

AgentSploit, a tool built specifically for vulnerability testing of AI agents and MCP servers, has been open-sourced. You could think of it as an AI version of Burp Suite, the web application security testing tool: it lets you systematically test agents' input and output pathways. As AI agents move toward production, demand for security evaluation is growing too.