Where is FaceMind headquartered?

FaceMind is headquartered in Shanghai, China. The company has a team of 11-50 people.

What does FaceMind do?

FaceMind operates in the Foundation Models / LLMs sector. A research-driven AI lab building parameter-efficient world models and LLMs, known for the Adam's Law frequency theory adopted by Anthropic and the Looped World Models (LoopWM) architecture. Building next-generation world models through parameter-efficient recurrent architectures rather than brute-force scaling, enabling 100x parameter efficiency and adaptive compute for long-horizon environment prediction.

When was FaceMind founded?

FaceMind was founded in 2023. It is based in Shanghai, China.

Who is the CEO of FaceMind?

Hongyuan Adam Lu (陆弘远) is the CEO of FaceMind. The company was founded in 2023.

Where is FaceMind located?

FaceMind is headquartered in Shanghai, China.

What funding stage is FaceMind at?

FaceMind is currently at the Growth Stage stage.

How many employees does FaceMind have?

FaceMind has approximately 11-50 employees.

FaceMind

Category: Foundation Models / LLMs

A research-driven AI lab building parameter-efficient world models and LLMs, known for the Adam's Law frequency theory adopted by Anthropic and the Looped World Models (LoopWM) architecture. FaceMind was founded in 2023. The company is led by Hongyuan Adam Lu (陆弘远). Based in Shanghai, China. Team size: 11-50. Total funding raised: $8.0M. Key investors include 星连资本 (Star Capital), 360 Group / 三六零 (Qihoo 360), 奇绩创坛 (MiraclePlus / Lu Qi's YC China).

Founded: 2023
Headquarters: Shanghai, China
Team size: 11-50
Total funding: $8.0M

Value proposition

Building next-generation world models through parameter-efficient recurrent architectures rather than brute-force scaling, enabling 100x parameter efficiency and adaptive compute for long-horizon environment prediction.

Products and solutions

1) Looped World Models (LoopWM) — parameter-efficient recurrent world model architecture with 100x parameter efficiency, 2) Adam's Law / Textual Frequency Law (TFL) — LLM frequency-aware framework adopted by Anthropic, 3) 叠叠社 (DieDieShe) — AI danmaku/overlay product for desktop/mobile/web, 4) Ego-NeuroLoop — human-centric closed-loop data for embodied AI, 5) animate.ai — AI virtual companion platform

Unique value

First to introduce Looped Transformer architecture to world modeling with shared-parameter recurrent blocks, achieving SOTA results at 1B scale while competitors require 100x more parameters; Adam's Law (text frequency theory) validated and adopted by Anthropic.

Target customer

Robot manufacturers (本体厂商), content platforms, chip & cloud vendors, GUI Agent developers, embodied AI companies, and C-end consumers via the 叠叠社 product.

Industries served

Robotics/Embodied AI, GUI Agents, Gaming/Entertainment (AI danmaku), AI Infrastructure (world models)

Technology advantage

1) Looped World Models (LoopWM) — shared-parameter recurrent Transformer blocks enabling iterative latent refinement with spectral stability constraints; 2) 100x parameter efficiency vs traditional world models; 3) Adam's Law (Textual Frequency Law) — foundational LLM theory adopted by Anthropic; 4) Deferred Decoding for efficient long-horizon rollout; 5) Early exit adaptive compute mechanism; 6) Ego-NeuroLoop for human-centric embodied data

How they differentiate

Unlike most world model efforts that scale parameters and data, FaceMind focuses on architectural innovation — using looped/recurrent shared-parameter blocks to achieve deeper computation without parameter bloat. Their 1B model matches or exceeds models with 100x more parameters on ScienceWorld benchmarks. They combine fundamental LLM research (Adam's Law) with practical product deployment (叠叠社).

Main competitors

1) 极佳视界 (Jijia Shijie) — Chinese world model startup backed by Huawei, 2) World Labs (Fei-Fei Li's spatial intelligence startup), 3) Google DeepMind (Genie, world model research)

Key partnerships

360 Group (strategic investor & partner), Anthropic (adopted Adam's Law research), NVIDIA (acknowledged LoopWM work), The Chinese University of Hong Kong (research collaboration), 深渡资本 (Shendu Capital, financial advisor)

Notable customers

C-end users of 叠叠社 (DieDieShe) AI danmaku product, early validation partners in embodied robotics, GUI Agent, and robotic arm environments

Major milestones

2023-06: Company founded, 2024-07: Angel round from 奇绩创坛 (MiraclePlus), 2025: Published SLoW paper on low-frequency words, 2025-06: Pre-A round from 360, 2026: Adam's Law paper (ACL 2026) adopted by Anthropic, 2026-06: Pre-A round led by Star Capital with 360 follow-on, 2026-06: Looped World Models paper released, #1 on HuggingFace Papers, 2026-06: NVIDIA acknowledged LoopWM work, 2026-07: Ego-NeuroLoop research published

Growth metrics

1B parameter LoopWM model achieves SOTA on ScienceWorld (68.4% EM, 85.3% F1) vs Claude-opus-4-6-max (47.2% EM, 72.8% F1); 100x parameter efficiency demonstrated

Market positioning

Early-stage Chinese AI research lab positioned at the intersection of world models and LLM fundamentals. Competing globally on architectural innovation rather than scale, with validation from Anthropic (Adam's Law adoption) and NVIDIA (LoopWM recognition).

Geographic focus

China (primary), with global research recognition (Anthropic adoption, HuggingFace #1, NVIDIA attention)

Patents and IP

Multiple ACL/NAACL/EMNLP publications; Adam's Law (ACL 2026 Main); Looped World Models (arXiv 2606.18208, HuggingFace Papers #1); Ego-NeuroLoop research; SLoW (low-frequency word impact on LLM translation, 2025)

About Hongyuan Adam Lu (陆弘远)

Ex-Microsoft Research Asia (MSRA) Pretraining team; Imperial College London CS BSc+MSc; University of Edinburgh AI MSc; PhD CUHK NLP Lab (Prof. Lin Wei). ACL Outstanding Paper Award first author; developed the first 200-language LLM at MSRA; ACL 2025 & NAACL 2025 Area Chair.

Official website: https://faceminda.com/