The optimal inference batch size for a sparse model like DeepSeek-V3 is more than 2,400 concurrent sequences.