Mistral Large 4 is 1.05T parameters with 52B active per token. The 52B sets your compute. The 1T sets your memory, because every expert has to stay resident in GPU memory for the router to pick it. At FP8 that is about a terabyte of weights before any KV cache.
That makes 8x B200 (1,440 GB) the first single node I'd plan FP8 on. Rented at about $53.52 an hour, it has to sustain roughly 7,100 output tokens per second all month to beat Mistral's preview API price, about 3,600 at list price. Most teams won't have that concurrency.
My read: the open weights matter most as an option that caps vendor leverage. Running them yourself pays off only for organisations that already run GPU clusters.
Memory tables, hardware tiers and the break-even chart: Mistral Large 4 is Open-Weight, But Not Easy to Run
Top comments (0)