Sapient押注小模型在推理上胜过LLM；开源HRM收获11.4K星标

做的是什么生意

Sapient Intelligence builds brain-inspired AI architectures, starting with the Hierarchical Reasoning Model (HRM) — a two-module recurrent system (slow planner plus fast executor) targeting complex, deterministic reasoning tasks where LLMs fail, with enterprise applications in robotics, logistics, healthcare and climate forecasting.

启动资金：$22M seed at a $200M valuation, led by Vertex Ventures, Sumitomo Group and JAFCO (announced December 2024).

起因

Founded in 2024 in Singapore by Guan Wang and Austin Zheng, Sapient Intelligence set out to challenge the assumption that better AI means bigger models. The team argued that autoregressive transformer LLMs are weak at long-horizon, multi-step reasoning, and turned to neuroscience: the brain organizes computation hierarchically across regions operating at different timescales (VentureBeat; Digit).

经过

In December 2024 the startup announced a $22M seed at a $200M valuation led by Vertex Ventures, Sumitomo Group and JAFCO (Digit). In mid-2025 it published HRM: two coupled recurrent modules — a slow high-level planner and a fast low-level executor — enabling 'hierarchical convergence' with 27M parameters trained on about 1,000 examples per task. VentureBeat reported an ARC-AGI score of 40.3% (vs 34.5% for o3-mini-high), near-perfect results on Sudoku-Extreme and Maze-Hard (where chain-of-thought models scored 0%), and a claimed ~100x speedup on task completion. The code was released on GitHub in August 2025 and became a ROSS Index Q4 2025 entry: 11.4K stars, 11.4x growth, as of 2025-10-24.

可借鉴之处

Sapient的楔子就是证据：一个在困难基准上胜过巨头的小型、开放、可复现的模型比任何论文都更有说服力——GitHub星标作为信誉引擎。