inclusionAI/Ling-2.5-1T · Hugging Face
This new 'instant model' claims to match frontier AI with 4x less compute.
Deep Dive
InclusionAI has open-sourced Ling-2.5-1T, a new flagship model with 1 trillion total parameters (63B active). It's trained on 29T tokens and processes up to 1M context length. The company claims its reasoning approaches frontier 'thinking models' while using roughly 4x fewer output tokens, due to a new composite reward mechanism and hybrid linear attention architecture. It also shows improved alignment for creative writing and tool-calling.
Why It Matters
If its efficiency claims hold, it could dramatically lower the cost and barrier to running high-performance AI.