Rather than impose one serving configuration everywhere, we tune quantization and speculative decoding for each platform.
1 more from this model
Its quantization and speculative-decoding configurations can be tuned to the target platform and workload.