A new technique lets smaller AI models absorb intelligence from larger ones by sharing mathematical weights, slashing costs and boosting performance. Mostik’s breakthrough could reshape how open-weight models compete with tech giants.
Mostik, a Russian startup, has developed a way for AI models to communicate directly by sharing their mathematical weights. This lets smaller models draw on the capabilities of much larger ones, without the usual back-and-forth of generating and passing text outputs. The result: faster, cheaper, and more efficient AI systems.
This method isn’t just a lab experiment. Mostik’s approach pushed their model to the top of the ARC-AGI 3 leaderboard, a well-known challenge in the AI field. The team, led by CEO Sasha Malysheva and chief scientist Stanislav Smirnov, hasn’t shared all the technical details yet, but their public demo is clear: they linked the 753-billion-parameter GLM-5.2 model with a much smaller, 4-billion-parameter Qwen-3.5 that can run on a mobile device. The combined system performs at a level exactly between the two, but at just one-twentieth the cost of running the full GLM model.
Stanislav Smirnov, Mostik's chief scientist, is a professor at the University of Geneva and a Fields Medal laureate (2010).
“It’s well-known in machine learning that ensembles outperform individual models,” Malysheva said. But Mostik’s technique skips the usual step of generating text, allowing models to share knowledge directly. This could make open-weight models much more competitive with closed systems from companies like Anthropic and OpenAI, which currently dominate the market with proprietary architectures.
According to WIRED, Mostik says its technology lets AI models exchange information through their weights, not just by swapping text responses. This direct approach is designed to cut computational costs while keeping performance high.
Stanislav Smirnov, a Fields Medalist and Mostik’s chief scientist, notes that finding a shared mathematical language for models to communicate is a major challenge. “There seems to be no appropriate mathematical language yet,” he said. Mostik’s solution is a literal bridge-one that could eventually help us understand how both AI and human brains tackle complex problems.
WIRED reports that if Mostik's approach is validated, it could strengthen open-weight models and help close the gap with closed systems from companies like Anthropic and OpenAI. The company asserts that its hybrid system is about 20 times cheaper than running the full GLM-5.2 model, while delivering quality that sits precisely between the large and small models.
For publishers and AI content creators, the impact is immediate. By letting smaller, cheaper models inherit abilities from larger ones, Mostik’s method could sharply reduce the cost of running advanced AI in production. Karl Tuyls, formerly of Google DeepMind, calls the method a “no-brainer” for anyone focused on efficiency. Vladimir Arustamian, tech lead at Lovable, thinks this could speed up the training of specialized models in fields like biology and physics, since domain-specific models could be paired with generalist models for faster results.
Malysheva’s own story is one of persistence. She was discouraged from pursuing advanced math early on, but went on to study at a top St. Petersburg school and now leads a team that has achieved in months what many thought would take years. “They said it might be too hard for a young girl,” she recalled. “I decided I need to prove them wrong.”
Mostik’s breakthrough isn’t just about technical efficiency. If open-weight models can be quickly upgraded and combined, the dominance of closed AI giants could face real competition. For content creators, this means more accessible and affordable AI tools are on the way. The next big step in AI may not come from building ever-larger models, but from connecting different ones to multiply their strengths.
As Superpower Daily reports, Mostik’s public demo showed that linking the massive GLM-5.2 model with the smaller Qwen-3.5 makes it possible to run advanced AI on mobile devices at a fraction of the usual cost.
Founded in 2024, Mostik is based in St. Petersburg and has fewer than 20 employees. Despite its small size, the startup has already caught the attention of major AI labs and academics. The company hasn’t shared funding details, but its rapid technical progress and competition results have made it a rising name in the open-weight AI world.
