Mistral Achieves '72% of Intelligence' With Just 3.8% of the GPUs, Sparking Debate

A viral post by @Totzenberger claims Mistral 4 achieves about 72% of frontier-model intelligence using only 3,800 Grace Blackwell GPUs—roughly 3.8% of the compute used by GPT-6 Astra—sparking debate over compute efficiency.

2026-10-07 ~ 2026-10-08 · 2 related posts

• "72% of the intelligence with 3.8% of the GPUs": Mistral's compute-efficiency ratio sparks debate — cyb3rops · 2026-10-07
• Mistral hits 72% of the intelligence with just 3.8% of the GPUs, fan math claims — JFPuget · 2026-10-08