Alexey Fateev(@superalesha):2.38 bpw and MRCR moves 0.2 points. That sounds like a big win. I want to check that number myself but 214GiB does not fit in 96GB, so the model is out of reach here. Their encoder change is not. It works on any model. @TencentHunyuan ship imatrix.gguf and the layer selection rule and I will run the same recipe on a MoE that fits on 4 3090s. I even have a couple of good candidates for this surgery.

2026.08.29 05:56

2.38 bpw and MRCR moves 0.2 points. That sounds like a big win. I want to check that number myself but 214GiB does not fit in 96GB, so the model is out of reach here. Their encoder change is not. It works on any model. @TencentHunyuan ship imatrix.gguf and the layer selection rule and I will run the same recipe on a MoE that fits on 4 3090s. I even have a couple of good candidates for this surgery.

显示更多

Tencent Hy@TencentHunyuan

2026.08.29 05:31

We compressed Hy4-preview from 1.5TB to ～200GiB GGUF and it still works well ! Meet MIX-STQ1_0.The trick isn’t just going low, it’s deciding where: calibration data picks each layer’s bit-width, some down to 1.31-bit STQ1_0, some up to 2.06-bit IQ2_XXS. Same budget, lower error. Accuracy barely moves vs BF16 📊 MCP Atlas 83.7→83.2 📊 SWE-Bench multi 82.9→81.3 📊 MRCR 81.3→81.1 📊 IFBench 73.5→72.5 See the details on HF : AngelSlim/Hy4-preview-GGUF Weights & low-bit GGUFs 👇 #LLM# #Quantization# #llamacpp# #Hy#

显示更多