Recently, Meituan officially released and open-sourced its latest AI large model - LongCat-Flash-Chat. This model demonstrates an excellent balance between computational efficiency and performance with a total parameter count of 560B and an activation parameter range of 18.6B to 31.3B. LongCat-Flash adopts an innovative mixture of experts (MoE) architecture, utilizing the "zero computation expert" mechanism, which activates only the necessary parameters for each token, ensuring efficient use of computing power.
In terms of model architecture, LongCat-Flash also introduces a cross-layer channel design, significantly improving the parallelism of training and inference. This allows the model to achieve a reasoning speed of 100 tokens per second for a single user within just 30 days of training on H800 hardware. Notably, during the training process, a PID controller was used to adjust the bias of experts in real-time, keeping the average number of activated parameters at 27B, effectively controlling computing power consumption.
Besides, LongCat-Flash has made many optimizations in enhancing agent capabilities. By building its own Agentic evaluation set and multi-agent data generation strategy, the model performs excellently in various agent tasks, especially in complex scenarios, where it ranked first in the VitaBench benchmark test. Compared to models with larger parameter scales, LongCat-Flash still demonstrates outstanding agent tool usage capabilities.
In terms of general knowledge, LongCat-Flash is no less impressive. In the ArenaHard-V2 test, it scored 86.50, ranking second among all evaluated models; and it achieved high scores of 89.71 and 90.44 in the MMLU and CEval benchmarks, showing its competitiveness in language understanding and Chinese ability assessment.
LongCat-Flash-Chat, with its efficient reasoning speed and outstanding agent performance, not only leads competitors technologically but also provides developers with more research and application opportunities through its open-source initiative.
Project Address: https://github.com/meituan-longcat/LongCat-Flash-Chat
Experience Website: https://longcat.ai/
