ISTA has released a coder variant of DeepSeek-V4.1-Flash-1017 compressed using GSQ and RCO quantization techniques. The sparse mixture-of-experts model originally features 384 routed experts per layer with 6 active per token, totaling 763.2B parameters. This release prunes half the experts, retaining 192 per layer, reducing transformer weights from 138.2 GB to 72.1 GB in GGUF format.