Russian startup Mostik bridges LLM hidden states, cutting cost to 1/20
机器之心 · wechat · 2026-09-06
Mostik, a little-known Russian startup with 12 PhDs and Fields medalist Stanislav Smirnov, briefly topped the ARC-AGI-3 Kaggle leaderboard after just four months, keeping its competition system secret.
Core idea: models exchange hidden states directly
- Today's multi-agent systems pass text, but a model outputs only 17 bits per token while its internal state holds 2MB — like thinking in a thousand dimensions and speaking through a keyhole.
- Mostik trains a "bridge" (the only trained component; both models stay frozen) that translates GLM-5.2's (753B params) hidden states into Qwen-3.5 (4B params), letting the small model handle generation.
Results
- The hybrid system costs 1/20 of running GLM alone, gains 25% accuracy, closes 50% of the capability gap, and up to 2x on hard tasks.
- Hidden-state handoff beats text handoff by up to 10 points everywhere tested, with the advantage largest the earlier the large model hands off.
Extensions
- A new interpretability entry point: compare sender/translated/receiver states and run intervention experiments to test causal features.
- Next: training-time distillation with live teacher guidance, and "specialization plug-ins" — betting on model-to-model collaboration as a third scaling route.
All results are from Mostik's own disclosures; no external reproduction yet.
Related event: Startup Mostik Tops ARC-AGI 3 with Latent-Space Model Bridge(3 posts)→
More from Models
- Diffusion LLM bias rankings shift with sampling configs, persist after debiasing — Bollegala · 2026-10-23
- Running ~100 DeepSeek v4.1 flash agents in parallel for under $20/month — zack_overflow · 2026-10-23
- Researcher laments open-source audio models are surprisingly subpar, even representation ones — gowthami_s · 2026-10-23
- User claims they generated video with "Claude Opus 5.5", unannounced model name surfaces — jwt0625 · 2026-10-23
- Security researchers push back as OpenAI restricts Astra access; BlackHat 2027 mocked as 'cyber model' bazaar — cyb3rops · 2026-10-23
- Mistral Large 4 Leaked on Hugging Face: 1T Params, 49B Active, Open Weights Oct 31 — mervenoyann · 2026-10-23