Standard multi-agent workflows still rely on sequential prompting, forcing neural networks to generate, serialize, and decode text tokens whenever they pass context. This token-based handoff creates crippling latency and multiplies token costs across production loops. AI startup Mostik—named after the Russian word for bridge—proposes bypassing textual tokenization entirely, allowing models to exchange internal hidden states and weight values directly.
Hybrid Architectures and Inference Costs
Developed by Russian mathematicians, Mostik's technique injects internal representations from a heavy donor model directly into a lightweight executor. In a recent demonstration, the team bridged two open-weight models: the 753-billion-parameter GLM-5.2 and a compact 4-billion-parameter Qwen-3.5 designed for edge deployment. The resulting hybrid ran at roughly one-twentieth the operational cost of the monolithic flagship while reclaiming half the performance delta between the two architectures.
As Sasha Malysheva, chief executive officer of Mostik, argued during an architecture briefing, chaining specialized weights fundamentally outpaces brute-force monolithic scaling:
"I personally do not think we will have a monolithic model [in the future] or that the capabilities of models will come from scaling,"
Malysheva contends that orchestrating heterogeneous networks via direct tensor transfers provides a practical alternative to the unsustainable compute overhead of ever-larger pre-trained monoliths.
Specialized Models and Benchmark Validation
Beyond raw inference savings, this non-verbal context transfer demonstrates clear reasoning gains over standard prompt-chaining and distillation. Applying the bridge method to the rigorous ARC-AGI 3 benchmark, Mostik's hybrid configuration climbed to the top of the leaderboard. The team has withheld full operational specs while actively competing for the top prize.
Industry observers point to major economic upside for enterprise agent pipelines. Vladimir Arustamian, tech lead at AI software company Lovable, noted that pairing frontier models with domain-specific architectures in biology or physics will accelerate fine-tuned model deployment. Arustamian highlighted that Mostik delivered a functioning latent-bridging implementation in months rather than years.
If non-verbal state transmission scales, direct hidden-state bridging gives open-weight ecosystems a viable architectural lever to outmaneuver proprietary API monopolies from OpenAI and Anthropic.