On April 22, Ant Group's Baichuan large model officially unveiled Ling-2.6-flash, an Instruct model boasting a total of 104 billion parameters and an active parameter count of 7.4 billion. This model places a strong emphasis on 'Token efficiency,' delivering swifter inference speeds and reduced costs without compromising on intelligence levels, thus rendering it a more fitting choice for extensive real-world applications. According to evaluations conducted by Artificial Analysis, Ling-2.6-flash accomplished tasks using 15 million tokens, consuming roughly one-tenth of the resources required by other models. In Agent-related benchmark tests, such as BFCL-V4 and TAU2-bench, the model achieved state-of-the-art (SOTA) performance within its size category. Currently, the API for Ling-2.6-flash is officially accessible, priced at $0.1 per million tokens for input and $0.3 for output, with a limited-time free trial spanning one week.