初创公司 Mostik 由俄罗斯数学家创立，开发出一种宛如「机器心灵感应」的技术，让不同的人工智慧模型能直接透过权重的数学数值进行沟通，而无须依赖耗时耗资的文字输出。这项突破性方法能将大型模型的能力高效转移至小型模型，不仅让团队在极具挑战性的 ARC-AGI 3 竞赛中迅速拔得头筹，更成功将 7530 亿参数的 GLM-5.2 与仅 40 亿参数的 Qwen-3.5 结合，以原模型二十分之一的成本实现了两者效能中位数的混合系统。
Mostik 执行长 Sasha Malysheva 指出，AI 的未来并非依赖单一巨型模型的单纯扩展，而更像是群体平均估计重量的智慧结合。这项技术使开源权重模型在面对 Anthropic 和 OpenAI 等封闭模型时更具竞争力。业界专家如 Lovable 的 Vladimir Arustamian 与前 Google DeepMind 电脑科学家 Karl Tuyls 皆高度赞赏此技术，认为它大幅提前了将尖端模型与物理、生物等特定专业领域模型高效配对的进程，是极致追求运算效率的理想方案。
Mostik 首席科学家兼 2010 年费尔兹奖得主 Stanislav Smirnov 表示，在不同 AI 模型间建立共通语言在数学上极为困难，而 Mostik 的架构成功搭建了这座桥梁，未来甚至有助于揭示 AI 与人类大脑在解决复杂问题时的共通推理机制。创办人 Malysheva 回顾自身成长历程与面对外界对年轻女性能力质疑的挑战，展现了证明技术可行性的强大决心，带领团队持续推进这项颠覆性的 AI 连接技术。
Startup Mostik, founded by Russian mathematicians, has developed a technique akin to 'machine telepathy' that allows different artificial intelligence models to communicate directly through their weight values without relying on text output. This approach efficiently transfers the capabilities of larger models to smaller ones, propelling the team to the top of the ARC-AGI 3 competition. They also demonstrated this by bridging the 753-billion-parameter GLM-5.2 with a 4-billion-parameter Qwen-3.5, achieving a hybrid system with performance halfway between the two at only one-twentieth the operating cost.
Mostik CEO Sasha Malysheva argues that the future of AI lies in combining diverse models rather than scaling monolithic architectures, analogous to the wisdom of crowds estimating weight. This direct communication method significantly enhances the competitive advantage of open-weight models against proprietary giants like OpenAI and Anthropic. Industry figures such as Lovable's Vladimir Arustamian and former Google DeepMind scientist Karl Tuyls praised the breakthrough for enabling small models to approximate large-model performance and accelerating the integration of frontier models with domain-specific systems.
Fields Medalist Stanislav Smirnov, serving as Mostik's chief scientist, noted that bridging two AI models mathematically addresses the lack of a common foundational language, potentially uncovering shared reasoning principles between artificial neural networks and the human brain. Driven by early mathematical talent and a determination to disprove skepticism regarding her youth and gender, Malysheva and her team continue to pioneer this innovative bridge technology, offering a new trajectory for efficient artificial intelligence collaboration. (Key numbers: 2010)