Developer Creates 'Fusion' Architecture: Fuses Models into a Sub-6B Sparse Powerhouse

menhguin · x · 2026-08-06

An independent developer released fuse-1-Lite, a new sub-6B sparse activation AI model built on a brand new architecture called Fusion. The model fuses the weights of LiquidAI's LFM2.5-2.6B and Alibaba's Qwen3.6-35B-A3B.

According to the creator, the model achieves near Qwen3.6-35B-A3B performance while being only a fifth of the size. The developer also plans to release 26B and 12B models based on this architecture, acting as mini versions of MiniMax and DeepSeek models, and is asking for compute donations to run benchmarks.

Original post →

More from Models

Models channel →