MOSTIK tops ARC-AGI leaderboard by piping frontier-model reasoning into small models via latent space
SimplyAnnisa · x · 2026-09-03
A team of 12 PhDs working for four months, MOSTIK claims first place on the ARC-AGI leaderboard. Its core idea: a frontier model only needs to do the reasoning — a small model on your own infrastructure converts it into the final answer.
- Models communicate directly in latent space: hidden states pass from the frontier model into the small one via a protocol, with no text in between
- No model-specific fine-tuning required, and different model families can be mixed
- If it scales reliably, this could dramatically cut compute costs without sacrificing reasoning quality
Competition details remain under wraps while the contest is still running.
Related event: Mostik bridges large and small models via latent space, tops ARC-AGI 3(3 posts)→
More from Infra
- llama.cpp deprecates --chat-template-kwargs, reasoning-preserve now on by default — Bulky-Priority6824 · 2026-09-03
- Agentic API adds a stateful layer in front of vLLM for open-model agent runtimes — techNmak · 2026-09-03
- Google's Gemini 3.8 Flash 'works harder' but may burn more tokens at same pricing — The Verge AI · 2026-09-03
- Mitchell Hashimoto Details Memory Optimization Tricks in the Superlogical Server — sull · 2026-09-03
- Perplexity's Lily beats MLX-LM with 1.23x prefill and 1.35x decode throughput on M5 Max — perplexity_ai · 2026-09-03
- Perplexity open-sources Lily, a local inference engine for Qwen3.6 on Apple silicon — perplexity_ai · 2026-09-03