CSTL, the team now sitting in first place on the ARC-AGI-3 Kaggle competition leaderboard, has been identified as Mostik, a newly public AI startup. Its current score is 7.51%.
The team has 15 members, including 12 PhDs and one Fields Medalist. The contest is not over yet. The second milestone closes on Sept. 30, while final submissions are due on Nov. 2.
Mostik is building a bridge between models
Mostik’s core idea is to let different AI models exchange internal states directly. Standard multi-agent systems mostly pass information through text. Mostik instead trains a bridge between two models, converting one model’s hidden states into a representation the other can read, without fine-tuning either model itself.
In a public experiment, the company used GLM-5.2 at 753B and Qwen3.5 at 4B. In simplified terms, the larger model handled prefill and the smaller model handled decode. GLM-5.2 read the problem but did not generate tokens. Its internal states were then passed straight to the 4B model, which produced the answer.
What the company says the method can do
Mostik said this approach can reduce the performance gap between the two models by about half. It also said that, compared with a medium-sized single model reaching the same accuracy, the compute cost is only about 40%. Even so, the final result still remains below the standalone 753B large model.
The company has not disclosed which model and harness were used to achieve the 7.51% ARC-AGI-3 result. The 753B-plus-4B setup described above was presented as a separate technical demo and should not be treated as the exact configuration behind the leaderboard result.
ARC-AGI-3 evaluates the full agent system
The report noted that ARC-AGI-3 is not simply a test of model size. It evaluates the full agent system. A previous stage winner used only Qwen 3.6 27B, yet still took first place through its harness, tools, and context management.
Limits and criticism
The method requires access to a model’s hidden states. That means cloud models such as Claude and GPT cannot currently be used in this setup.
Google DeepMind engineer Susan Zhang also questioned the engineering tradeoff. If developers can already control a model’s internal states, she asked whether freezing two models and then training a separate bridge is really the most cost-effective engineering path.