LongCat-Flash-Chat

2025-09-01 · Publisher : Meituan · Domain : Artificial Intelligence

Meituan opens LongCat-Flash-Chat, a mixture-of-experts model with 560 billion parameters, of which 18.6 to 31.3 billion are active depending on the context, under the MIT license, very fast at inference and strong on agent tasks.

Meituan opens LongCat-Flash-Chat, a mixture-of-experts model with 560 billion parameters, of which 18.6 to 31.3 billion are active depending on the context, under the MIT license, very fast at inference and strong on agent tasks. ### Context In September 2025, the Chinese open ecosystem already includes DeepSeek-V3.1, Qwen3 and Kimi-K2. Meituan, a delivery and local services platform, joins the movement with a model whose design targets throughput: the technical report states training on more than 20 trillion tokens in 30 days. ### What the model brings - Zero-computation experts: a mechanism that allocates a dynamic compute budget to important tokens, which makes the active parameters vary from 18.6 to 31.3 billion (about 27 billion on average); a PID controller on the expert bias maintains the average. - Shortcut-connected MoE (ScMoE): widens the overlap window between computation and communication, for a throughput above 100 tokens per second. - Stability at scale: hyperparameter transfer from small models, initialization by growth from a checkpoint half the size, deterministic computation and detection of silent data corruption. - Agent training: two-stage pretraining concentrating the reasoning data, mid-training up to a 128,000-token context, and task synthesis by a multi-agent framework along three axes (information processing, tool complexity, user interaction). - Non-reasoning model: LongCat-Flash-Chat is the chat version; a Thinking variant followed (not covered here). ### Architecture and specifications - Architecture: MoE with 560 billion parameters in total, 18.6 to 31.3 billion active (average of about 27 billion). - Context: 128,000 tokens. Training: more than 20 trillion tokens in 30 days. Modalities: text. - License: MIT. Weights: meituan-longcat/LongCat-Flash-Chat on Hugging Face; technical report on arXiv. - Knowledge cutoff: not disclosed. ### Cost and access The technical report states a serving cost of 0.70 dollars per million output tokens at more than 100 tokens per second; this is an estimate by Meituan for its own deployment, not a verified API price. Self-hosting hardware is not specified in the model card; LMSYS published a post on deployment with SGLang. | Item | Information | | --- | --- | | Estimated serving cost (technical report) | 0.70 dollars per million output tokens | | Official API pricing | Not verified | | License | MIT | ### Benchmarks | Benchmark | LongCat-Flash | GPT-4.1 | Claude 4 Sonnet | Gemini 2.5 Flash | DeepSeek V3.1 (open) | | --- | --- | --- | --- | --- | --- | | MMLU | 89.71 | 89.64 | 91.75 | 86.33 | 90.96 | | ArenaHard-V2 | 86.50 | 61.50 | 62.10 | 77.00 | 84.10 | | AIME 2024 (avg@10) | 70.42 | 47.00 | 47.00 | 79.67 | 66.30* | | GPQA-Diamond | 73.23 | 67.68 | 70.71 | 80.30 | 74.90* | | LiveCodeBench (pass@1) | 48.02 | 39.21 | 45.59 | 39.65 | 56.40* | | SWE-Bench-Verified | 60.40 | 48.60 | 68.00* | 40.60 | 66.00* | | TerminalBench | 39.51 | 28.40 | 40.74 | 12.35 | 31.30* | | τ²-Bench (telecom) | 73.68 | 35.20 | 46.20 | 16.50 | 38.50 | | IFEval | 89.65 | 85.58 | 88.35 | 83.92 | 86.69 | | GraphWalks-128k | 51.05 | 85.02 | 80.57 | 64.83 | 73.54 | Source: Hugging Face model card of LongCat-Flash-Chat. Scores measured by Meituan, single table; *: value taken from a public report. Claude 4 Sonnet, Gemini 2.5 Flash and DeepSeek V3.1 are evaluated in non-reasoning mode. ### Versus closed models In the vendor's table, LongCat-Flash outperforms GPT-4.1 on all the lines retained except GraphWalks-128k (51.05 vs 85.02), with MMLU being a tie, and is ahead of Claude 4 Sonnet on ArenaHard-V2 (86.50 vs 62.10), AIME 2024 (70.42 vs 47.00), IFEval and above all τ²-Bench telecom (73.68 vs 46.20). It is behind Claude 4 Sonnet in software engineering (SWE-Bench-Verified 60.40 vs 68.00) and only slightly behind on TerminalBench (39.51 vs 40.74), and behind Gemini 2.5 Flash on GPQA-Diamond (73.23 vs 80.30) and AIME 2024 (70.42 vs 79.67). Long context is a clear weak point: 51.05 on GraphWalks-128k. The closed models compared are GPT-4.1, Claude 4 Sonnet and Gemini 2.5 Flash in non-reasoning mode, mid-range 2025 models: no closed reasoning model is in the table. The price gap is not established for lack of a verified API price. ### Key takeaways LongCat-Flash-Chat is Meituan's entry milestone: a 560-billion-parameter MoE with dynamic compute, very fast, whose results from the vendor often exceed GPT-4.1 and approach Claude 4 Sonnet on agents. Its MIT license and detailed technical report make it a reference for efficient design.

Tags : LLM, Open Source, Meituan, LongCat, MIT, MoE, Agents