This isn't unreasonable. ML workloads benefit from more computational time per request. Lower QPS = better results.