The archive · AI & Models · Technical decision · 2024–2026
Sapient bets tiny models beat LLMs at reasoning; open-source HRM hits 11.4K stars
Singapore's Sapient Intelligence raised $22M at a $200M valuation to prove a 27M-parameter hierarchical reasoning model can outthink LLMs — open-sourced it.
Sapient Intelligence
What the business is
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.
Starting capital：$22M seed at a $200M valuation, led by Vertex Ventures, Sumitomo Group and JAFCO (announced December 2024).
How it started
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).
What happened
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.
No ending yet — it is still running.
Background
Sapient Intelligence is a Singapore-based AI startup founded in 2024 by Guan Wang and Austin Zheng. Its thesis is contrarian: instead of scaling parameters, build an architecture inspired by the brain, with a slow high-level module for abstract planning and a fast low-level module for detailed computation. The result, the Hierarchical Reasoning Model (HRM), uses only 27 million parameters and about 1,000 training examples per task — no large pre-training corpus, no chain-of-thought data.
The company raised a $22M seed at a $200M valuation in December 2024, led by Vertex Ventures, Sumitomo Group and JAFCO, to pursue this alternative to transformer scaling. When HRM was published in mid-2025, VentureBeat reported an ARC-AGI score of 40.3% — ahead of o3-mini-high (34.5%) and Claude 3.7 Sonnet (21.2%) — near-perfect performance on extreme Sudoku and hard maze tasks where chain-of-thought models scored 0%, and a claimed ~100x speedup in task-completion time for deterministic problems.
Sapient open-sourced HRM in August 2025. The bet on reproducible evidence worked as distribution: sapientinc/HRM hit 11.4K GitHub stars with 11.4x growth by 24 October 2025, making ROSS Index Q4 2025's list of the fastest-growing open-source startups, and Forbes named co-founder Wang Guan to 30 Under 30 Asia. The company is now working to evolve HRM into a more general-purpose reasoning module, citing early results in healthcare, climate forecasting and robotics.
The wager, in one line: for a class of hard reasoning problems, a smarter, smaller, structured architecture can beat the giants — and the quickest way to make that claim believable was to give the evidence away.
What has to be true
- Autoregressive LLMs are weak at long-horizon multi-step reasoning, so Sapient targeted a real gap rather than competing head-on in language tasks.
- A 27M-parameter model with 1,000 training examples is a falsifiable, reproducible claim — open-sourcing it converted benchmark results into community trust and GitHub stars.
- The $200M seed valuation rested on architecture, not product: investors bet that brain-inspired hierarchical computation could become a general reasoning layer rather than another model race.
- By targeting deterministic, latency-sensitive domains (robotics, logistics, diagnostics), Sapient picked problems where 'smarter, smaller and faster' has a clear economic value.
What can be applied
Sapient's wedge was evidence: a small, open, reproducible model that beats giants on hard benchmarks is more convincing than any paper — GitHub stars as the credibility engine.
Aftermath
As of 02 September 2026 Sapient Intelligence is live and seed-funded: HRM remains open source on GitHub (11.4K stars by late October 2025, ROSS Index Q4 2025), and the company says it is evolving HRM from a specialized problem-solver into a more general-purpose reasoning module, with early work in healthcare, climate forecasting and robotics — including a reported 97% accuracy on seasonal climate forecasts (Korben, 2025-08-27).
Sources
- New AI architecture delivers 100x faster reasoning than LLMs with just 1,000 training examples
- Sapient's RNN AI model aims to surpass ChatGPT and Gemini: Here's how
- HRM - The 27M AI that crushes GPT-4 on reasoning
- Top trending open-source startups in Q4 2025
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card