Pathway · 2026-08-10 · notable
BDH-CQ — a 150M model that reasons in latent space on ARC-AGI-1
BDH-CQ is a 150M-parameter model from Pathway that scores 29.5% pass@2 on the public ARC-AGI-1 evaluation set at $0.00070 per task. It reasons inside a recurrent latent state instead of writing out chain-of-thought text.
Pathway's BDH-CQ solves ARC-AGI-1 puzzles by thinking in a recurrent latent state instead of writing out a chain of thought.
Key specs
| Parameters | 150M |
|---|---|
| Arc agi 1 pass@2 | 29.5% |
| Cost per task | $0.00070 |
What is it?
A 150M-parameter reasoning model called BDH-CQ scores 29.5% pass@2 on the public ARC-AGI-1 evaluation set at a computed $0.00070 per task. Pathway compares that with GPT 5.6 Luna (Low), which scores 34.2% at $0.040 per task. The weights are not released, and Pathway says the exact update rules and implementation details stay proprietary.
How does it work?
Demonstrations shown at inference time update a recurrent memory one step at a time, so what a later example contributes depends on associations built from earlier ones. The model then runs a set number of iterative steps inside a high-dimensional latent workspace and decodes an answer from the final state. Nothing is written out as text, so BDH-CQ leaves no intermediate reasoning trace.
Why does it matter?
Cost per solved task is the claim worth watching. Pathway reports one H200 GPU takes about 0.85 seconds per task at standard effort, and had the run reproduced independently by Łukasz Kaiser, a co-author of the 2017 Transformer paper. If latent reasoning holds up at larger sizes, small models could take on puzzle-style reasoning that today needs a frontier model and a long chain of thought.
Who is it for?
ML researchers
Try it
git clone https://github.com/pathwaycom/arc-task-gen