Google is quietly winning the open-weights model race

Gemma 4 12B dropped as a unified, encoder-free multimodal model, and the HN thread flagged this as a significant architectural move. The encoder-free design is unusual enough that even enthusiasts in the thread admit they don't fully understand it yet, which is a good sign that it's genuinely novel rather than a repackaging.

The pattern here: Google is releasing capable open-weights models at a pace that's making the community compare them favorably to where Meta was before the Llama 4 release. Several commenters note that Google's business case for open releases is unclear, which is exactly what makes it dangerous to competitors. When a company with Google's resources gives away competitive infrastructure, it compresses the moat for everyone charging for API access to similar capability.

The LLM security thread adds texture: someone tested multiple models on a vulnerable app and found Anthropic models underperformed not because of capability, but because their guardrails blocked the attacks. Open weights models don't have those guardrails, which is a double-edged sword for builders.