Unlocking 560B‑Parameter AI: Inside LongCat‑Flash‑Chat’s Zero‑Computation MoE

Technical Highlights

LongCat‑Flash uses a Mixture‑of‑Experts architecture with Zero‑Computation Experts. Total parameters are 560 B, but only 18.6‑31.3 B are activated per token (average 27 B), controlled by a PID controller to keep activation stable.

Cross‑layer channels enable parallel MoE communication, achieving fast training (30 days) and inference speed over 100 tokens/s on H800. Training incorporates hyper‑parameter transfer, model stacking, and stability strategies.

For agentic capabilities, a dedicated evaluation set and multi‑agent data generation improve performance.

Performance Evaluation

Comprehensive benchmarks show LongCat‑Flash matches or exceeds leading models despite smaller size. In ArenaHard‑V2 it scores 86.50 (2nd), MMLU 89.71, CEval 90.44. In agentic tasks it leads τ2‑Bench and ranks first on VitaBench (24.30). In programming benchmarks it scores 39.51 on TerminalBench (2nd) and 60.4 on SWE‑Bench‑Verified. In instruction following it tops IFEval with 89.65 and achieves best scores on COLLIE and Meeseeks‑zh.

Model Deployment

Two efficient deployment scripts are provided for SGLang and vLLM. Example for single‑node SGLang:

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

Further deployment instructions are in the GitHub repository.

Open Source

The model, code, and weights are released under the MIT License and can be used for downstream tasks, model distillation, etc. Access the model at https://longcat.ai/, Hugging Face, and GitHub links.