LongCat-Flash-Chat: 560B MoE Model with 27B Active Params Sets New Benchmarks

Technical Highlights

LongCat-Flash adopts an innovative Mixture‑of‑Experts (MoE) architecture called Zero‑Computation Experts. The model contains a total of 560 B parameters, but each token activates only 18.6 B–31.3 B (average 27 B) parameters, achieving on‑demand compute allocation. A PID controller fine‑tunes expert bias during training to keep the average activation around 27 B.

The model introduces cross‑layer channels that parallelize MoE communication and computation, dramatically improving training and inference efficiency. Combined with custom low‑level optimizations, LongCat‑Flash completes training within 30 days and achieves inference speeds of over 100 tokens / s per user on H800 GPUs.

Training incorporates hyper‑parameter transfer and model stacking, along with multiple strategies to ensure stability, resulting in a smooth and efficient training process.

For agentic capabilities, LongCat‑Flash built a dedicated Agentic evaluation set and optimized the entire training pipeline, including multi‑agent generated diverse high‑quality trajectory data, leading to superior agent performance.

Performance Evaluation

General Knowledge : Achieves 86.50 on ArenaHard‑V2 (2nd place), 89.71 on MMLU, and 90.44 on CEval, rivaling leading domestic models despite a smaller parameter count.

Agentic Tool Use : Outperforms larger models on τ2‑Bench and ranks first with a score of 24.30 on VitaBench for complex scenarios.

Programming : Scores 39.51 on TerminalBench (2nd place) and 60.4 on SWE‑Bench‑Verified, demonstrating strong command‑line and software engineering abilities.

Instruction Following : Leads with 89.65 on IFEval and achieves top results on COLLIE (57.10) and Meeseeks‑zh (43.03), showing excellent compliance with complex multilingual instructions.

Model Deployment

Two efficient deployment solutions are provided based on SGLang and vLLM, enabling quick setup and experimentation.

python3 -m sglang.launch_server \
    --model meituan-longcat/LongCat-Flash-Chat-FP8 \
    --trust-remote-code \
    --attention-backend flashinfer \
    --enable-ep-moe \
    --tp 8

For detailed deployment instructions, refer to the LongCat‑Flash‑Chat repository.

Visit https://longcat.ai/ to start a conversation with LongCat‑Flash‑Chat.

Open Access

The source code is released under the MIT License, allowing users to leverage model outputs, perform model distillation, and train derivative models.

Hugging Face: https://huggingface.co/meituan-longcat/LongCat-Flash-Chat

GitHub: https://github.com/meituan-longcat/LongCat-Flash-Chat