Meituan Unveils the Groundbreaking LongCat-Flash-Chat: An Open-Source AI Model
On September 1st, Meituan made a significant announcement with the release of its innovative AI model, LongCat-Flash-Chat, alongside its open-source framework. This advancement marks a pivotal moment in the AI domain, showcasing impressive capabilities and a fresh architectural approach.
An Overview of LongCat-Flash-Chat
LongCat-Flash-Chat leverages a cutting-edge Mixture-of-Experts (MoE) architecture. This model boasts a staggering 560 billion total parameters, with an activation range of 18.6 billion to 31.3 billion parameters depending on context. This unique setup allows for optimal computational efficiency and enhanced performance, distinguishing LongCat-Flash-Chat from existing models.
The Zero-Computation Experts Mechanism
At the heart of LongCat-Flash-Chat is the "Zero-Computation Experts" mechanism. With this architecture, only the necessary parameters are activated based on the context of each token, enabling a more efficient allocation of computational resources. To regulate power consumption effectively during training, a PID controller is deployed, ensuring that the average active capacity for a single token remains around 27 billion parameters.
Innovative Features for Enhanced Performance
The LongCat-Flash model introduces cross-layer channels facilitating MoE communication, allowing for extensive parallel processing. This breakthrough significantly boosts both training and inference efficiency. The model’s training is completed in just 30 days, achieving an impressive inference speed of over 100 tokens per second for a single user on the H800 platform.
LongCat-Flash-Chat also refines commonly utilized large model components and training methodologies. Through hyperparameter migration and model layer superposition, the team has successfully ensured stable and reliable training outcomes.
Superior Agentic Capabilities
LongCat-Flash-Chat has integrated a comprehensive Agentic evaluation set to refine data strategies. This model employs multi-agent techniques, generating diversified and high-quality data trajectories, thus enhancing its agentic abilities.
Through a thoughtful combination of algorithmic design and engineering optimizations, LongCat-Flash-Chat achieves theoretical costs and speeds that surpass industry standards. It boasts a generation speed of 100 tokens per second on the H800, with output costs as low as RMB 5 per million tokens.
Exceptional Benchmark Performance
A thorough evaluation against multiple benchmarks indicates that LongCat-Flash-Chat competes remarkably well against the leading models in the market. Particularly, it excels in agent-related tasks while swiftly activating only a fraction of its parameters.
General Knowledge Acumen
In the ArenaHard-V2 benchmark, LongCat-Flash-Chat scored 86.50, securing the second position among all evaluated models. It also delivered outstanding results in the MMLU (Multi-task Language Understanding) benchmark with a score of 89.71, and achieved 90.44 in the CEval (Chinese General Capability Assessment), demonstrating its capability in general domain knowledge.
Competence in Agentic Tool Use
In tests for agent tool usage, LongCat-Flash-Chat maintained a competitive edge. Even when compared to models with a greater number of parameters, it outperformed in the τ2-Bench, and ranked first in complex scenario tasks with a score of 24.30 in VitaBench.
Programming Skills
When evaluated on the terminalBench (focused on command line tasks), LongCat-Flash-Chat garnered a score of 39.51, securing the second position. Additionally, it showcased expertise in the SWE-Bench-Verified, targeted at software engineering capabilities.
Instruction Compliance
The model also excelled in following instructions, scoring 89.65 in the IFEval benchmark and achieving top marks in other evaluations, confirming its proficiency in language command sets across both English and Chinese.
Conclusion
Currently, LongCat-Flash-Chat is available as open-source software on both GitHub and Hugging Face platforms, enabling developers and researchers to explore its capabilities. The release not only contributes to the vast landscape of artificial intelligence but also sets a new benchmark for future developments in AI technologies.
This momentous launch positions Meituan at the forefront of AI innovation, offering a robust solution that combines deep learning with operational efficiency and versatility. As the AI landscape continues to evolve, LongCat-Flash-Chat stands ready to push boundaries, inviting further exploration and application across various domains.
