Model Style not-a-Bench
I've been disliking the recent trend of LLMs releasing with very promising and high benchmarks, only to disappoint by failing simple tasks or sanity checks. Plus, benchmarks don't really tell you about the LLMs' stylistic choices. So I made 'Model Style not-a-Bench' to be able to quickly eyeball various LLMs, most preferably at 4-bit quants, under 100GB weights (excluding storage offloads like Qwen Flash Next's approach). https://msnab.pages.dev