@MistralAI Engineering + astronomy on a private STEMBench is a nicer signal than another public leaderboard reshuffle. How much of Large 4's gain is tool use vs raw knowledge—any public split you can share?
@reflection_ai Apache 2.0 plus BF16/FP8/NVFP4 out of the gate is the right packaging—most open releases make you hunt for a usable stack. How does DFlash2 draft quality hold up in multi-tool agent loops vs full BF16?