This paper proposes a reconfigurable Single Flux Quantum (SFQ) accumulator architecture with multi-precision support, specifically optimized for energy-efficient neural inference in superconducting computing systems. By leveraging hierarchical Toggle Flip-Flop (TFF) cascading and asynchronous pulse-driven logic, the design dynamically adapts to varying operand bit-widths. The modular architecture employs a hybrid of TFF-based state propagation and clockless carry merging, enabling scalable accumulation for multi-operand sequences. With configurable output truncation and adaptive clock gating, the design supports continuous data streaming at up to 142.8 GHz.