InclusionAI Releases Ling-3.1-flash-1019 Weights and Base Model
InclusionAI has released the open weights for Ling-3.1-flash-1019 on Hugging Face, including both the chat variant and the Ling-3.1-flash-base model in BF16 format. This update corresponds to improved benchmark performance, such as a rise to 44 on the AA Intelligence Index v4.3.2 and 68% on AutomationBench-AA. The release provides the local LLM and AI open-source community with a large-scale Mixture of Experts model featuring a 1 million context window. Furthermore, making the base model available in BF16 allows researchers and developers to conduct their own post-training rather than relying solely on pre-finished chat models. The model retains a 560B total parameter size with 25B active parameters and a 1M context window. Users should note that the SGLang serving example in the 1019 README requires updating the model path to inclusionAI/Ling-3.1-flash-1019.
## BACKGROUND
Mixture of Experts is an architectural design where only a subset of the network's parameters is activated for any given input, enabling high model capacity with lower computational costs during inference. BF16 is a 16-bit floating-point numerical format commonly used in machine learning to balance training performance and memory efficiency.