Mostik Claims Top of ARC-AGI 3 with Latent-Space Model Bridging

Startup Mostik (Russian for "bridge") claims its hybrid approach tops the ARC-AGI 3 leaderboard and has been covered by WIRED. Instead of asking whether open models can catch up with frontier models, the team questions why final answers must come from the frontier model at all: a 753B frontier model handles reading and reasoning, then passes hidden states directly in latent space—via a protocol, without textual handoff or fine-tuning—to a 4B model that writes the answer. @udmrzn notes the competition is still ongoing and the team is keeping technical details secret.

Confirmed

• Mostik claims the top of ARC-AGI 3 (relayed by multiple posters; @nordicinst notes WIRED has covered the company)
• Technical route: frontier-model reasoning + latent-space hidden-state transfer to a 4B model, no text handoff, no fine-tuning
• Team background: posts including @SimplyAnnisa say 12 PhDs over four months; investor ashvardanian (relayed by @viglovikov) describes founders from Google X, NVIDIA, Perplexity and FAR.AI, with 15 researchers and engineers
• Claimed gains: 1/20 cost (@SimplyAnnisa), 2.5x less compute (@udmrzn), 20x faster inference (@TheMoonMidas)

Unconfirmed

• Performance figures (20x faster, 1/20 cost, 2.5x compute savings) come from the team or secondhand accounts with no independent verification
• Team size differs across sources (12 PhDs vs 15 researchers/engineers)

Why it matters

• The approach offers a new paradigm beyond textual handoff in multi-model systems: direct weight/latent-space communication—what @nordicinst calls "machine telepathy"—injecting frontier capabilities into small models more efficiently than ensembles
• If latent-space communication is reproducible, it could upend the narrative that open models must chase the frontier alone, sharply cutting the cost of intelligent reasoning

2026-09-03 ~ 2026-09-04 · 7 related posts

• [source] Startup Mostik bridges AI models via their weights, tops ARC-AGI 3 at 1/20 the cost — nordicinst · 2026-09-03
• [source] Mostik bridges frontier and small models in latent space, tops ARC-AGI 3 at 1/20th the cost — SimplyAnnisa · 2026-09-03
• MOSTIK tops ARC-AGI leaderboard by piping frontier-model reasoning into small models via latent space — SimplyAnnisa · 2026-09-03
• Mostik's latent bridge lets a 4B model read a 753B model's mind, halving the gap at 2.5x less compute — udmrzn · 2026-09-03
• 753B model thinks, 4B model writes: latent-space handoff matches frontier reasoning at 20x speed — TheMoonMidas · 2026-09-03
• [source] Ex-Google X, NVIDIA Staff Found mostik: Models Talking in Latent Space, ARC-AGI #1 Claim — viglovikov · 2026-09-03
• Mostik's 12-PhD team tops ARC-AGI leaderboard by passing hidden states from frontier to small models — VoidAsuka · 2026-09-04