Open weights don't make a 1T model cheap to run

Mistral Large 4 has 1.05T parameters with 52B active per token: the 52B sets compute, while the 1T sets memory because every expert must stay resident in GPU memory. At FP8 that is about a terabyte of weights, making 8x B200 (1,440 GB) the first single node for FP8 deployment at roughly $53.52 an hour.

• Mistral Large 4: 1.05T parameters, 52B active per token
• At FP8 weights take about 1 TB before any KV cache
• Deployment needs an 8x B200 node with 1,440 GB at about $53.52/hour
• Self-hosting pays off only for teams already running GPU clusters