A new open-source model named Laya has been developed, designed for the Jev architecture. This model, with 421 million parameters, utilizes a non-autoregressive decision model approach and was trained on a large, human-annotated corpus of over 25,000 real-world examples. The training was conducted on a single RTX 6000 Pro with 96 GB of VRAM, and the model is capable of running on lower-end PCs due to its small size. Laya reportedly surpasses existing Jev benchmarks and includes an RLCD policy-gradient reinforcement learning approach for optimized decision-making. AI
IMPACT Provides a new, efficient open-source model for specific architectures, potentially enabling wider adoption on less powerful hardware.
RANK_REASON Release of an open-source model with benchmark results. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
