50. They're comparing a pre-RL base checkpoint to fully post-trained models without disclosing the massive inference… by @DJLougen (Daniel Lougen, M.S.) · backlist 2026-05-08 · rubric 91.0