🤖 Russian Mostik Teaches AI Models to Communicate via Weights Instead of Text

The startup connected the open 753-billion parameter GLM-5.2 with the mobile 4-billion Qwen-3.5: the hybrid is 20 times cheaper than the large model, and the quality is right in the middle between them. WIRED calls this 'machine telepathy'. A model that took first place in the ARC-AGI 3 competition is already being built — details are kept secret.

🌍 If 'bridges' between models scale, progress may come not from a monolith, but from a network of specialized models: the quality of the large model is transferred to the small one through weights. This reduces inference costs and increases the value of open models, but for now this is just a demo figure from the company itself.

👤 Keep an eye on the ARC-AGI 3 leaderboard: details are promised to be revealed after the competition. Quality may work even on a phone — cheap local hybrids instead of expensive APIs.

Source 1: https://www.wired.com/story/russian-startup-mostik-ai-models-communication/