BDH-CQ
#12 in Reasoning ModelsPathway · 2× · last seen Aug 12, 2026
33
Momentum
BDH-CQ is a 150-million-parameter Reasoning model from Pathway, an AI lab that developed a post-Transformer architecture called "Dragon Hatchling" (BDH). Instead of generating Chain-of-Thought as text like classical Transformers, BDH-CQ solves tasks in a recurrent latent state, which according to the provider enables significantly lower Inference costs. On the public ARC-AGI-1 Benchmark, the model achieved 29.5% pass@2 at a calculated cost of $0.0007 per task, which Pathway claims is approximately 11 times cheaper than GPT 5.6 Luna (Low) with only slightly lower accuracy. The underlying BDH architecture is available as code on GitHub.
Momentum trend
15.05.13.08.
Features
| Key Benchmark (%) | 29.5% pass@2 on ARC-AGI-1 (public eval set) |
| License | BDH architecture code available as open-source reference implementation on GitHub (pathwaycom/bdh); BDH-CQ itself not licensed as a standalone release |
| Platform | Proprietary post-transformer architecture (Dragon Hatchling/BDH), trained using Amazon SageMaker HyperPod |
| Price per 1M Tokens | No per-token price given; computed inference cost of $0.0007 per ARC-AGI-1 task |
| Release Date | August 11, 2026 |