Ant Group's Bailing Large Model Launches Ling-2.6-flash: Anonymous Test Version Topped OpenRouter Trending List

Ant Group's Bailing large model recently launched Ling-2.6-flash, an Instruct model with 10.4 billion total parameters and 7.4 billion active parameters. Previously released anonymously as Elephant Alpha on OpenRouter, it saw continuous growth in usage, topping the Trending list for multiple days with a daily token call volume of around 100 billion.

Ling-2.6-flash optimizes inference efficiency via a hybrid linear architecture, reaching up to 340 tokens per second on 4 H20 cards. It calibrates token efficiency during training, consuming only 1/10 of the tokens of similar models in evaluations. It also enhances tool calling and task execution capabilities for Agent scenarios, performing well in multiple benchmark tests.

Currently, the model offers a one-week free API call on OpenRouter and the official platform. After the free period, it will be charged by usage. The official also stated that versions like BF16, FP8, and INT4 will be open-sourced soon.