6 articles

The Chinese AI startup's new flagship processes vastly longer prompts while undercutting Western competitors on price.

New model closes gap with frontier AI systems while improving efficiency over previous versions.

KV cache and tokenizer bugs squashed. Local inference actually viable now.

Google's open model sweeps comparisons. Early users report better reasoning and lower memory demands than expected.

Google DeepMind's new Gemma 4 family spans 1B to 31B parameters with native multimodal support and 256K context.

The industry is moving beyond either-or thinking. Diverse AI architectures will power every company, every country, and every app.