Confirmed
ISTA releases expert-pruned DeepSeek-V4.1-Flash-1017 GSQ-RCO Coder at about 1 bit per parameter
The Coder build keeps 192 of 384 routed experts per layer, reducing transformer weights to 72.1 GB from 138.2 GB. Reported retention is 98.7% on LiveCodeBench v6 (88.57 vs 89.71) and 94.8% on SWE-bench Verified (77.00 vs 81.20). It runs in llama.cpp b11870 or newer and in Strata.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead
[Release] ISTA DeepSeek-V4.1-Flash-1017 GSQ-RCO Coder: a 763B model at 1 bit per parameter