Quantization after SFT and RL

Quantization after SFT and RL

Neither SFT checkpoint showed a clear NVFP4 penalty

MXFP8NVFP4No change

Qwen 3.8 27B on an internal multi-turn reasoning dataset, 800 held-out examples. Each quantized checkpoint is compared with its own BF16 version, so positive values mean reward decreased. Intervals are 95% paired, difficulty-stratified bootstraps that cluster the two GRPO rollouts per example.