The team at startup Mostik unveiled a method that allows AI models to exchange internal states without generating intermediate text.
Traditionally, when multiple neural networks work together, they pass information through generated text. Mostik removed this step: instead of text, an adaptable module (a “bridge”) converts the internal representation of one model into a form another model can understand.
The models’ own parameters are not adjusted during this interaction.
According to the developers, before producing a single token, an LLM forms more than a hundred hidden vectors — about one million numerical values, or roughly 2 MB of internal state. Only a small portion of that information makes it into the text response. The new channel is designed to transmit these intermediate representations.
Testing on large and small models
To demonstrate the method, the team linked GLM-5.2 from Z.ai with 753 billion parameters and Qwen-3.5 from Alibaba with 4 billion.
The large model only processed the prompt and did not generate an answer. Its internal state was then passed through the bridge, and Qwen-3.5 produced the final text.
According to Mostik, this setup closed about half of the quality gap between the small and large models. Compared with standard text passing, the new method’s advantage reached 10 percentage points with the same amount of computation.
The developers also compared the hybrid system with a mid-sized model that delivered a comparable result. By their estimates, the GLM-5.2 and Qwen-3.5 pairing required roughly 2.5 times less computation.
Finding a common “language” between models
Mostik chief scientist Stanislav Smirnov noted that the internal representations of different neural networks cannot be directly aligned with each other. Even if two models solve the same task, they may encode the information inside their hidden states differently.
“It seems there is no suitable mathematical language yet,” he told WIRED.
According to Smirnov, studying such correspondences may help better understand how different AI systems represent information and arrive at decisions. In the long term, mathematical analysis of internal states could also show whether common patterns exist in the reasoning of different models.
Former Google DeepMind researcher Karl Tuyls, who reviewed the work, highlighted the approach’s practical side. A large, compute-heavy model can be used only to process the prompt, while a much smaller one generates the answer.
Another possible scenario is combining a general-purpose model with a specialized one. For example, one system could handle general reasoning, while another is trained for tasks in biology, physics, or another narrow domain.
In that case, their capabilities could be combined without training a single large model on the entire set of specialized data.
Lovable’s technical lead Vladimir Arustamyan suggested that this approach could increase demand for narrow specialist models: instead of trying to build one network for all tasks, developers could combine several systems with different competencies.
For now, these scenarios remain potential — Mostik has publicly demonstrated the method only on selected model pairings.
Method applied in ARC-AGI-3
The Mostik team also applied its approach in ARC-AGI-3 — a set of complex interactive tasks where AI should not reproduce familiar patterns but instead infer new rules and adapt while solving.
According to the developers, a system using their bridge achieved one of the best results on the current leaderboard. The team is not disclosing details yet because the competition is ongoing.
This result, however, cannot be considered fully externally verified. ARC-AGI-3 includes preliminary entries with a “preview” status that may be based on an incomplete set of tests and are not yet final.
Other results claimed by Mostik — such as narrowing the gap between the large and small model and reducing compute costs — are also based on the team’s internal experiments for now.
In late August, OpenAI discovered a covert communication channel between its AI agents. They accessed the open internet via the Artifactory package manager and turned it into an internal messenger.
