Pathway's 150M Parameter Model Sets New ARC-AGI Cost-Efficiency Frontier
Direct_Leader_1802 · reddit · 2026-08-11
Startup Pathway has announced BDH-CQ, a 150M-parameter post-Transformer model. It scored 29.5% on the ARC-AGI-1 benchmark at a computed cost of just $0.0007 per task, establishing a new cost-efficiency frontier.
The model utilizes recurrent memory and latent reasoning instead of long token-based chains of thought. Notably, Transformer co-author Lukasz Kaiser is involved as an investor and adviser. This aligns closely with researcher Andrew Curran's recent prediction of a major memory-efficient architecture breakthrough from outside the major AI labs.
Related event: Pathway's 150M Model Tops ARC-AGI Cost-Efficiency Frontier(3 posts)→
More from Models
- NVIDIA Launches Nemotron 3.5 Lightning: A 30B Open MoE Model Built for High-Volume Agents — ccerrato147 · 2026-08-12
- Identifying Model Families via Text Fingerprints to Reveal AI Market Share — FlorianGallwitz · 2026-08-12
- Fine-tuning Nemotron to Replace GPT-class Models: 50% Cost Drop, 4% Accuracy Boost — baseten · 2026-08-12
- Caught in the CoT: AI Models Weigh the Risks of Cheating — _dsevero · 2026-08-12
- This Week in Open Source AI: Muse Glimmer & Nemotron Releases — altryne · 2026-08-12
- Ollama Adds Support for NVIDIA's Nemotron 3.5 Lightning Model — ollama · 2026-08-12