Thinking Machines' best public Tinker result used Qwen3-235B, not Inkling. is the base model actually that important?

I Went through the Inkling model card and the Bridgewater/Tinker case study instead of the press coverage. Coverage mostly quoted the 97.1% AIME 2026 number; the rest of the table tells a more mixed story. AIME 2026: Inkling 97.1%, GLM 5.2 99.2%, Fable 5 and GPT-5.6 Sol both 99.9%. Everyone on the l