逆矩阵科技完成超亿美元种子++轮融资，发布通用世界基座模型Physis-v0.1
英文摘要
World model startup Physis (逆矩阵科技) has closed a seed++ funding round surpassing $100 million, led by Matrix Partners China, 5Y Capital, and光合创投, with strategic investment from Ant Group and continued support from existing investors Gaocheng Capital and Yan Yuan Capital. The company simultaneously released Physis-v0.1, a general world foundation model that aims to serve embodied intelligence, industrial simulation, game physics, and scientific prediction through one pre-training run, focusing on physical correctness, long-range consistency, causal action, and generalization. Physis states the window for a major leap in world base model capabilities has compressed to 18 months, and plans a flagship model by end of 2026 with open-source slices and technical reports. The team, founded by Peking University scholars Chen Boyuan and Ji Jiaming, combines young researchers and experienced engineers in a flat, AI-native organization, and argues that scaling in a "physical latent space"—not pixel space—is key to unlocking a scaling law distinct from language models.
中文摘要
世界模型公司逆矩阵科技已完成超亿美元种子++轮融资，本轮由经纬创投、五源资本、光合创投等领投，蚂蚁集团战略投资，高瓴创投、燕缘创投持续加注。公司同步发布通用世界基座模型Physis-v0.1，定位“One For All”，一次预训练即可服务具身智能、工业仿真、游戏物理及科学预测等多场景，主打物理正确、长程一致、动作因果和通用泛化能力。逆矩阵认为通用世界基座模型的窗口期已压缩至18个月，计划2026年底推出旗舰模型并放出开源切片与技术报告。团队由北大青年学者陈博远、吉嘉铭联合创立，融合奥赛金牌得主与企业资深工程人才，构建AI-native扁平化组织，并强调需在物理隐空间进行Scaling以实现不同于语言模型的指数级预测误差下降。
关键要点
Physis raises over $100 million in a seed++ round from leading VCs and Ant Group, following a previous multi-million-dollar round just two months prior.
逆矩阵科技完成超亿美元种子++轮融资，由经纬创投、五源资本等领投，蚂蚁集团战略参投，距上轮融资仅两月。
The company releases Physis-v0.1, a universal world foundation model designed to serve multiple physical-world applications from a single pre-training, emphasizing physical correctness, causality, long-range consistency, and generalization.
发布通用世界基座模型Physis-v0.1，通过一次预训练覆盖具身智能、工业仿真、游戏物理等场景，核心能力为物理正确、动作因果、长程一致及通用泛化。
Physis asserts that the window for a major capability leap in world base models has shrunk to 18 months, and full multi-scenario deployment is expected within 36 months, analogous to the GPT-3 to ChatGPT path.
逆矩阵判断世界基座模型能力跃迁窗口期已压缩至18个月，36个月内多场景实际落地，路径对标大语言模型从GPT-3到ChatGPT。
The team is pursuing a scaling law in a physical latent space rather than pixel space, using reinforcement learning with verifiable physical rewards to move from correlation to causality and avoid visual redundancy.
技术路径上，团队在物理隐空间进行Scaling，通过强化学习及可验证物理奖励从相关性转向因果性，压缩视觉冗余。
A flagship model is planned for end of 2026, with open-source slices and technical reports; the company prioritizes building the base model over early commercialization, expecting that a strong foundation will naturally lead to vertical adaptation.
计划2026年底发布旗舰模型并放出开源切片，现阶段专注基座能力，不急于垂类变现，认为先通用再适配才能构建跨场景壁垒。