Research & Benchmarks

We present a novel calibration-aware quantization pipeline that achieves sub-2% perplexity degradation on INT4-quantized transformer models running on Apple A17 Pro and Snapdragon 8 Gen 4. Our approach combines activation-aware weight quantization (AWQ) with recursive distillation feedback loops, enabling deployment of 7B-class intelligence on devices with <6GB unified memory.

We introduce Recursive Feedback Distillation (RFD), a multi-cycle teacher-student training paradigm where the student model is iteratively evaluated against a 405B-parameter oracle across 1,200+ domain-specific benchmarks. After 47 refinement cycles, our 1B-parameter student achieves 97.3% of the teacher's performance on medical NLU tasks while requiring 99.7% less compute at inference time.

We explore the extreme quantization frontier by applying ternary weight representations ({-1, 0, +1}) to transformer architectures optimized for tactical edge environments. Our 1.58-bit models achieve 89.4% of FP16 accuracy on classification tasks while enabling inference on FPGA-based hardware with power budgets under 5W — critical for disconnected, denied, and degraded (D3) operational contexts.