BDH-CQ: A 150M Post-Transformer Model That Cuts Reasoning Costs by 11x

Pathway AI has published a preprint detailing BDH-CQ, a 150 million-parameter post-transformer cognition model that uses vector memories and a latent reasoning engine instead of traditional text-based chain-of-thought. On the ARC-AGI-1 benchmark, BDH-CQ scores nearly 30% and costs roughly 11 times less per token than OpenAI's GPT-5.6 Luna. For AI builders, this architecture suggests that reasoning efficiency may improve dramatically without requiring massive parameter counts or expensive inference.

What Makes BDH-CQ Different from Transformers

Standard transformer models store prompts and interaction histories as long strings of numerical values representing text. As reasoning chains grow, memory usage expands quadratically, creating bottlenecks in GPU memory. BDH-CQ replaces this with a post-transformer memory scheme that uses numerical arrays (vectors) to represent the underlying rules and relationships of a task. These vectors point to locations in a theoretical map stored in GPU memory, allowing the model to process abstract reasoning without expanding its memory footprint.

The model's latent reasoning engine works through iterative recurrent loops. It takes the output of the last loop, assesses how to improve it based on training data, and feeds the result back as the starting point for the next iteration. More loops increase thinking time, but memory and compute consumption remain constant. In theory, running 200 loops costs the same as running 20. Transformers, by contrast, burn exponentially more GPU memory for each additional reasoning step.

Benchmark Performance and Cost Comparison

BDH-CQ scored nearly 30% on ARC-AGI-1, solving three out of ten nonverbal reasoning puzzles in two or fewer attempts. While OpenAI's GPT-5.6 Luna (Low) achieved a slightly higher score, the study states that this "modest accuracy gain" cost roughly 11 times as much in relative token costs. BDH-CQ was trained on just 150 million parameters, compared to the tens or hundreds of billions used by frontier models like Meta's Llama 3 70B or 405B.

Independent Verification and Scaling Plans

The ARC-AGI-1 results have been independently reproduced by NYU researcher Richard Zhong and Lukasz Kaiser, a co-author of the original "Attention Is All You Need" paper. Kaiser stated that "Pathway shows that model architecture, not just scale, can drive the next leap in AI reasoning." Pathway plans to scale the BDH architecture up to 600 billion parameters and target ARC-AGI-2 and ARC-AGI-3 benchmarks, as well as develop a full LLM for applications in cybersecurity incident response and industrial operations.

What This Means for AI Builders

BDH-CQ is still a research preprint, and its real-world performance on diverse tasks beyond ARC-AGI puzzles remains unproven. The cost comparison is based on reported token costs during benchmarking, not production API pricing. But the architectural insight is significant: if vector-based reasoning can scale to larger parameter counts without the memory overhead of transformers, it could change the economics of running reasoning-heavy AI workloads. Builders working on agent loops, multi-step reasoning, or cost-sensitive inference should watch this line of research closely.