better context model (0.2-1% lower bitrate)