POCKET is a 34.66-billion-parameter sparse Mixture-of-Experts model engineered for efficient on-device inference by activating only about 3 billion parameters per token. The architecture uses domain-specific expert pruning, retaining only the language-relevant experts out of 256 total while preserving performance, combined with MoE-aware mixed-precision quantization that keeps shared computational paths at high precision while aggressively quantizing routing experts. This lets the model run on standard tools like llama.cpp on commodity hardware without custom forks or a GPU.
