LongCat-Flash-Chat
A 560-billion-parameter Mixture-of-Experts language model with dynamic computation, averaging 27B activated parameters per token. Built for high-throughput chat and agentic tasks.
Independent third-party resource. Not affiliated with or endorsed by LongCat or Meituan.
Overview
LongCat-Flash-Chat is a 560-billion-parameter Mixture-of-Experts (MoE) language model released by Meituan's LongCat team in August 2025. It was the first model in the LongCat family and introduced the Shortcut-connected MoE (ScMoE) architecture and Zero-Computation Experts mechanism that would later be scaled up in LongCat-2.0. On average, 27 billion parameters are activated per token (dynamic range 18.6B–31.3B), controlled by a PID feedback controller that adjusts expert bias in real time. The model was trained on over 20 trillion tokens in approximately 30 days and achieves inference throughput above 100 tokens per second on H800 GPUs. It supports a 256K context window (upgraded from 128K in December 2025) and multilingual output across nine languages. LongCat-Flash-Chat is released under the MIT License and is available via the LongCat API platform, HuggingFace, and self-hosted deployment.
Architecture
LongCat-Flash-Chat introduced two architectural innovations documented in the LongCat-Flash Technical Report (arXiv:2509.01322). The Shortcut-connected MoE (ScMoE) design places shortcut connections between MoE layers to enlarge the computation-communication overlap window, reducing the idle time that GPUs spend waiting for expert routing results. This proved critical for scaling to tens-of-thousands of accelerators. The Zero-Computation Experts mechanism introduces a dummy expert that handles simple tokens at zero compute cost — a PID controller dynamically tunes the bias of this dummy expert to maintain the target average activation rate of approximately 27 billion parameters per token. When facing computationally demanding inputs (complex reasoning, tool use), more real experts activate; for simple inputs (greetings, trivial completions), the model conserves compute. The publisher describes this as the first implementation of token-level dynamic compute allocation in an open-source MoE model.
Training
Training was completed on over 20 trillion tokens within approximately 30 days, using a cluster of H800 GPUs. The PID controller was implemented during training to stabilize the average activation rate at approximately 27B parameters per token. The publisher reports that this training efficiency was achieved at a cost of $0.70 per million output tokens. The model was upgraded in December 2025 with the context window expansion from 128K to 256K and additional multilingual support for nine languages.
Key Benchmarks
Publisher-reported scores. Not independently verified by this site. Descriptions summarize what each benchmark measures.
| Benchmark | Score | What It Measures |
|---|---|---|
| MMLU | 89.71 | Massive Multitask Language Understanding — 57-subject knowledge benchmark spanning STEM, humanities, and social sciences |
| MMLU-Pro | 82.68 | More challenging variant of MMLU with harder distractor options designed to resist superficial pattern matching |
| IFEval | 89.65 | Instruction Following Evaluation — measures ability to follow explicit formatting, length, and content constraints |
| MATH 500 | 96.40 | 500 competition-level mathematics problems requiring multi-step reasoning and formal solutions |
| Arena-Hard v2 | 86.50 | Human preference benchmark using GPT-4 as judge — measures response quality on challenging open-ended prompts |
| SWE-bench Verified | 60.40 | Software engineering benchmark — measures ability to generate correct patches for real GitHub issues |
| C-Eval | 90.44 | Chinese comprehensive evaluation benchmark covering 52 subjects across humanities, science, and engineering |
Deployment
LongCat-Flash-Chat is distributed on HuggingFace (meituan-longcat/LongCat-Flash-Chat) in BF16 and FP8 precision. The FP8 variant (meituan-longcat/LongCat-Flash-Chat-FP8) requires at least one 8-GPU node (e.g., 8x H20) for inference via SGLang or vLLM with tensor and expert parallelism. The BF16 variant requires at least two 8-GPU nodes (e.g., 16x H800). The LongCat API platform provides access with a 256K maximum output limit and 50,000,000 free tokens per day. The model can also be loaded via Hugging Face Transformers with trust_remote_code=True for smaller-scale experimentation, though full context utilization requires the recommended GPU configurations.
Capabilities
Described in the cited primary documentation. Not an independent performance evaluation.
- General conversational AI with 560B total parameters (MoE).
- Per-token dynamic activation of 18.6B–31.3B parameters (averaging ~27B).
- 256K context window (upgraded December 2025 from 128K).
- Strong agentic tool use performance: τ²-Bench 73.68, VitaBench 24.30.
- Instruction following: IFEval 89.65, COLLIE 57.10.
- 100+ tokens per second inference speed on H800 GPUs.
- Deployable via SGLang and vLLM; FP8 and BF16 precision variants available.
- Multilingual support across 9 languages including Spanish, French, Arabic, Portuguese, and Russian.
Access & License
MIT License. Permits commercial use, modification, redistribution, and model distillation.
Documented context window: 256,000 tokens (per the cited publisher API documentation).
Sources
- LongCat-Flash Technical Report (arXiv:2509.01322)
Primary sourcePublished 2025-09-01Accessed 2026-07-26
Technical report detailing the ScMoE architecture, Zero-Computation Experts mechanism, and training framework. Reports inference at >100 TPS on H800 and a training cost of $0.70 per million output tokens.
- LongCat-Flash-Chat GitHub Repository
Primary sourceAccessed 2026-07-26
Primary repository under MIT License. Contains model configuration, chat template, and deployment instructions for SGLang and vLLM.
- LongCat-Flash-Chat on HuggingFace
Primary sourceAccessed 2026-07-26
Model weights in BF16 and FP8 precision. Model card includes publisher-reported benchmark scores across general domains, coding, math, and agentic tool use.
- Meituan Official Announcement for LongCat-Flash-Chat
Publisher documentationPublished 2025-08-29Accessed 2026-07-26
Publisher press release confirming MIT License, H800 inference benchmark of >100 TPS, PID controller for expert bias, and output cost of approximately 5 RMB per million tokens.
Related guides
- LongCat API Quick Start
A primary-source-based introduction to creating a LongCat API key and sending a basic chat completion request through the OpenAI-compatible endpoint.
FAQ
- Where can the primary source for LongCat-Flash-Chat be found?
- The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-Flash-Chat.
- Is LongCat-Flash-Chat available through the LongCat API?
- LongCat-Flash-Chat is documented in the meituan-longcat GitHub organization. Refer to the publisher API documentation at longcat.chat for current API availability.
- Is this page an official LongCat or Meituan page?
- No. This page is published by an independent third-party site. It is not affiliated with, endorsed by, or sponsored by LongCat or Meituan.
- How often is this page updated?
- This page was last verified on 2026-07-26. Content is reviewed when new publisher documentation or model releases become available.
Independent third-party disclosure
This page is published by an independent third-party site. It is not affiliated with, endorsed by, sponsored by, or operated by Meituan, LongCat, or any of their affiliates. The content summarizes publicly-available primary documentation and does not represent the views of any referenced organization.
Last reviewed: 2026-07-26