Link 3.1 Flash
A large model series released by Ant Group. The Flash version emphasizes lightweight and fast inference.
Timeline (last 90 days)
Ant has released the Link 3.1 Flash weights, using a hybrid architecture of linear and full attention; the model card lists 560 billion total parameters and 25 billion activated parameters, with the license marked as MIT.
The Link 3.1 Flash model card states 1 million context, while the current configuration and deployment documentation have 256K verified, with deployment including 8 H200s or 4 GB300s, plus speculative decoding support.
StepFun's Step 5 Preview and Ant's Link 3.1 Flash both use a mixture-of-experts (MoE) architecture, where total parameters determine the model's knowledge capacity and activated parameters determine inference compute cost.
All Facts (3)
The Link 3.1 Flash model card states 1 million context, while the current configuration and deployment documentation have 256K verified, with deployment including 8 H200s or 4 GB300s, plus speculative decoding support.
50%UnverifiedStepFun's Step 5 Preview and Ant's Link 3.1 Flash both use a mixture-of-experts (MoE) architecture, where total parameters determine the model's knowledge capacity and activated parameters determine inference compute cost.
50%UnverifiedAnt has released the Link 3.1 Flash weights, using a hybrid architecture of linear and full attention; the model card lists 560 billion total parameters and 25 billion activated parameters, with the license marked as MIT.
50%