Hacker News · AI· sha_rad·· 3 hours agoSelectedAI score65
Dory1.0-4B Elastic-Depth Recurrent LLM Released with Open Checkpoints
Dory1.0-4B: A Looped Normalized-GPT
AI guide
A research team has open-sourced Dory1.0-4B, a 4-billion-parameter elastic-depth recurrent large language model that allows a single checkpoint to adapt to varying compute budgets during inference by adjusting the recurrence count of its shared modules.
Why it matters
By looping shared layers in latent space, the model achieves elastic-depth inference and delivers significant performance gains on long-context and agent tasks while caching only the final recurrence's KV cache.
Source:Hacker News · AI · huggingface.co