All the mainstream news is covered by major labs' models. But what goes unreported is the work by community on getting these to run on much lower hardware and much faster.
Many of us know about llama.cpp/sglang/vllm et al.
Still fewer know about
@turboderp_ 's exl3 and exllamav3,
@ashxhart 's Tensorfold, and now
@mmastrac @coffeedev @CK2084 's work on exlr8 and these are just a few of these examples which are playing as big a role as any in AI.