Mostik bridges frontier and small models in latent space, tops ARC-AGI 3 at 1/20th the cost
SimplyAnnisa · x · 2026-09-03
Startup Mostik, built by 12 mathematicians in four months, has topped the ARC-AGI 3 leaderboard using a novel approach: a frontier model does the reasoning, then hidden states are passed directly in latent space to a small model running on your own infra—no text in between, no fine-tuning. In a demo, they bridged the 753B-parameter GLM-5.2 with a 4B-parameter Qwen-3.5: the hybrid costs 1/20th of the full GLM model and performs exactly halfway between the two. CEO Sasha Malysheva argues the real question isn't whether open models catch up to frontier models, but why a frontier model should generate your answer at all when only its reasoning is needed.
Related event: Mostik Bridges Large and Small Models via Latent Space, Tops ARC-AGI 3(2 posts)→
More from Models
- How Could Chinese Open-Source Models Actually Hurt the US? Two Failure Modes Explained — matanSF · 2026-09-03
- Meta's Muse Spark 1.3 tops Gemini 3.8 Flash on most overlapping benchmarks, crushes long-context MRCR — ChrisGPT · 2026-09-03
- Insider Praises Gemini 3.8 Flash: Better at Requirements, More Critical, Fast — prajdabre · 2026-09-03
- Tested GLM-5.3 abliterated model: 4x the cost, worse performance than the original — BLUECOW009 · 2026-09-03
- Gemini 3.8 Flash Accused of Benchmark Overfitting, Regressing vs 3.7 in Independent Tests — bindureddy · 2026-09-03
- Submission timelines hint at long-horizon post-training: open models submit >10 hours late — nrehiew_ · 2026-09-03