Local ModelsMedium importance

ISTA researchers squeeze DeepSeek V4.1 Flash down to about 1 bit per parameter in a coding version

The DASLab at the Institute of Science and Technology Austria (ISTA) published a “Coder” version of DeepSeek-V4.1-Flash-1017 in GGUF, the format used by llama.cpp and LM Studio. They cut half of the experts (192 of 384 per layer remained) and compressed the rest with two methods of their own, GSQ and RCO; the transformer weights dropped to 72.1 GB, versus 138.2 GB for the unpruned version, which comes to about 1 bit per parameter of the original model. According to the authors, the model keeps 98.7% of the original's score on LiveCodeBench v6 (88.57 vs. 89.71) and 94.8% on SWE-bench Verified (77.00 vs. 81.20). A shared 110.6 GB file (Engram) is still required.