Google Gemma 2 27B will be released in June

They've been making slow but steady progress, though. Bard was clown shoes at launch, and never really got *good*, but it did steadily improve to the point where it wasn't a complete embarrassment. Gemini was a big step up from that, and gemini pro right now trades blows with GPT4 in certain tasks. You got to remember, there's a bit of a desert for mid-size hobbyist models. There's command-R 35b, but it lacks GQA so it's basically unusable if you want more than maybe 2-4k context. There are a few Chinese models like Yi, internLM, and Qwen, but they're very... Chinese. Weird and janky implementation and trained so heavily on Chinese data that their English performance suffers quite a bit. Yi has proved more or less impossible to tame, even with extensive finetuning. Nobody wants to risk throwing more effort into Chinese models when the return is likely to be the same. Currently, there's *nothing new or interesting* between ~7b models at the low end and 70b models at the high end except mixtral - which still won't fit into 24GB of VRAM without resorting to a small, crappy quant. So when a big western company who has been doing some decent (if not spectacular) LLM work lately says they're about to publish weights for a mid-sized model that will fit into 24GB easily or even 16GB at a squeeze, people are right to get excited about it. Maybe it falls flat on its face, but realistically, what else does anybody have to look forward to in this category?