AI가 움직이고 미래가 업데이트됩니다.Qwen3.8-27B-iMatrix-NVFP4-MTP-GGUF
I built this quant because the ready-made FP4 file answered the wrong question. It was fast, but on my short WikiText-2 control it scored 6.4949 PPL. Plain Q4_0 scored 6.3798. The first higher-quality hybrid went too far the other way: good perplexity, 34.19 tok/s, and no comfortable room for 256K plus vision. It is a 17.1 GB, 5.01 BPW mixed-precision GGUF of Qwen/Qwen3.8-27B. It keeps large, tolerant matrices in native NVFP4 and spends more bits on selected attention, Gated DeltaNet, and late FFN tensors. The trained MTP layer remains embedded in the same GGUF. This is not a fine-tune. I built the private calibration workload from 5,472 messages across 296 real sessions in my Hermes agent setup. The 153,600 processed tokens covered coding, infrastructure work, tool calls, and mixed Polish-English conversation. I used `llama-imatrix` to rank tensor sensitivity; the corpus and raw importance matrix are not distributed or embedded in the model.
번역 준비 중입니다. 원문을 표시합니다.
I built this quant because the ready-made FP4 file answered the wrong question. It was fast, but on my short WikiText-2 control it scored 6.4949 PPL. Plain Q4_0 scored 6.3798. The first higher-quality hybrid went too far the other way: good perplexity, 34.19 tok/s, and no comfortable room for 256K plus vision. It is a 17.1 GB, 5.01 BPW mixed-precision GGUF of Qwen/Qwen3.8-27B. It keeps large, tolerant matrices in native NVFP4 and spends more bits on selected attention, Gated DeltaNet, and late FFN tensors. The trained MTP layer remains embedded in the same GGUF. This is not a fine-tune. I built the private calibration workload from 5,472 messages across 296 real sessions in my Hermes agent setup. The 153,600 processed tokens covered coding, infrastructure work, tool calls, and mixed Polish-English conversation. I used `llama-imatrix` to rank tensor sensitivity; the corpus and raw importance matrix are not distributed or embedded in the model.