Calibration-Free Int4 Robustness in a Byte-Level Conversational Model

Low-bit quantization is important for offline and edge language models, but 4-bit inferenceoften requires calibration, specialized quantizers, or careful treatment of sensitive weight tables.This paper reports an empirical result from Voxyne, a small byte-level conversational model witha Rahu-Ketu σK control channel. On the full WikiText-2 raw test corpus, the released Voxyneidentity v2 checkpoint changes from 1.5443 to 1.5615 bits per byte under naive calibration-free int4 quantization, a 1.11% increase relative to its own full-precision baseline. The samebenchmark shows much larger relative degradation for token models under the same naive int4path. In a 12-prompt conversational coherence probe, Voxyne remains stable at naive int4, whilea SmolLM2-135M-Instruct baseline degenerates into repetition under naive int4 and recoversonly with GPTQ. The result does not claim superior absolute language quality. It shows thatthis byte-level design is unusually robust to naive int4 quantization under the reported loss andgeneration checks.