Mini toolsHuihui 4 8B A4B v2
By HuiHui ai
Huihui4-8B-A4B-v2 is a consumer-hardware-friendly MoE chat model built by pruning and tuning Google’s Gemma-4-26B-A4B-it. The model card says it reduces the original 128 experts to 32 while keeping 8 active experts per token, and uses supervised fine-tuning on huihui-ai/GLM-5.1-Multilingual-STEM data to better reflect GLM-5.1-style thinking mode. It is positioned for daily conversation, code assistance, lightweight task planning, and local deployment, with stated VRAM needs below 18 GB in FP16 and roughly 6 to 9 GB when quantized.
New
Multimodal
Overview
Huihui4-8B-A4B-v2 is a lightweight pruned MoE conversational model derived from Gemma-4-26B-A4B-it. It is designed for consumer-grade hardware and code-related conversation, using 32 experts with 8 active per token and adopting GLM-5.1-style thinking-mode data during supervised fine-tuning.
Last updated: July 7, 2026
