BDH-CQ Model: Latent Reasoning Slashes Inference Cost to $0.0007 per Task

TheTuringPost · x · 2026-08-17

Researchers introduced BDH-CQ, a 150M-parameter model that combines "latent reasoning" and "temporary memory" to achieve extremely low inference costs. Instead of writing out intermediate steps, the model processes its internal state within a hidden representation (latent space) and can learn rules directly from demonstration examples. On the ARC-AGI-1 visual reasoning benchmark, BDH-CQ achieved a 29.5% pass@2 rate with an inference cost of just $0.0007 per task. This suggests that language may not be a necessary workspace for AI reasoning, offering a new scaling path for small models.

Related event: Pathway's Tiny Model Sets ARC-AGI Cost-Efficiency Record(2 posts)→

Original post →

More from Research

Research channel →