Mostik Leads the ARC-AGI-3 Kaggle Competition with an AI Model Bridging Technology

ME AI News, Dongcha Beating AI Brief: CSTL, currently ranked first in the ARC-AGI-3 Kaggle competition, is the newly unveiled AI startup Mostik, with a current score of 7.51%. The team consists of 15 members, including 12 PhDs and one Fields Medalist. However, the competition is not over—the second milestone deadline is September 30, and final submissions are due on November 2. Mostik enables direct exchange of internal states between different AI models. While conventional multi-agent systems rely on text-based communication, Mostik trains a bridge between two models to convert one model’s hidden states into a representation the other can understand—without fine-tuning either model. In their publicly disclosed experiment, they used GLM-5.2 (753B) and Qwen3.5 (4B), which can be roughly understood as the large model handling prefill and the small model handling decode: GLM-5.2 reads the question but does not generate tokens; its internal state is directly passed to the 4B model to generate the answer. Mostik claims this approach reduces the performance gap between the two models by approximately half and requires only about 40% of the compute needed for a single medium-sized model achieving the same accuracy. However, the final result still falls short of the standalone 753B model. Mostik has not yet disclosed which specific model and harness produced the 7.51% score; the 753B + 4B setup mentioned above is merely a separate technical demonstration and should not be assumed as their competition configuration. The ARC-AGI-3 competition evaluates the entire agent system: previous stage winners achieved first place using only Qwen 3.6 27B, by leveraging sophisticated harnesses, tools, and context management. On another note, this technique requires access to model hidden states, making it incompatible with cloud-based models like Claude and GPT. Google DeepMind engineer Susan Zhang has also questioned whether, given the ability to control internal states, freezing two models and training a separate “bridge” is truly the most cost-effective engineering solution. (Source: BlockBeats)