Researchers at the Institute of Science and Technology Austria released a heavily compressed build of the 763-billion-parameter DeepSeek-V4.1-Flash model. The build prunes half of all experts and shrinks core weights to 72.1 gigabytes while retaining most coding performance.