LongCat-Flash-Thinking

A 560B-parameter open-source MoE reasoning model with domain-parallel training, DORA asynchronous RL, and 64.5% token reduction on AIME-25.

Independent third-party resource. Not affiliated with or endorsed by LongCat or Meituan.

Overview

LongCat-Flash-Thinking is a reasoning-optimized variant of the 560-billion-parameter LongCat-Flash MoE model family, released in September 2025. Unlike standard chat models that generate responses in a single pass, Flash-Thinking produces an extended Chain-of-Thought — reasoning through problems step by step before delivering the final answer. The publisher reports an outstanding result: reducing token consumption by 64.5% on AIME-25 compared to standard chain-of-thought approaches, without degrading accuracy. The upgraded Flash-Thinking-2601 variant (January 2026) introduces 'Heavy Thinking' mode that simultaneously explores 8 parallel reasoning paths, then synthesizes the most promising results. Flash-Thinking achieves a perfect score of 100.0 on AIME-25 (with heavy thinking) and 86.8 on IMO-AnswerBench, the current open-source state-of-the-art on that benchmark. It ranks first on BrowseComp (73.1) among all models, indicating strong agentic search capabilities.

Architecture

The core architectural innovation is domain-parallel training: instead of training a single model on a mixed dataset of STEM, code, and agentic tasks, the training process decouples optimization across these domains in parallel, then fuses the resulting expert models into a single nearly Pareto-optimal model. The DORA (Dynamic ORchestration for Asynchronous rollout) reinforcement learning framework supports this by coordinating training across more than 10,000 environments spanning over 20 domains, with a reported threefold training speedup over synchronous RL methods. The Heavy Thinking mode in Flash-Thinking-2601 activates 8 parallel reasoning chains simultaneously, explores divergent solution strategies, then consolidates the results into a final answer — a design the publisher describes as 're-thinking' rather than incremental refinement.

Training

The domain-parallel training scheme uses DORA to coordinate asynchronous RL rollouts across domain-specific environments. Each domain (STEM, Code, Agentic) trains under its own reward signal, then the experts are fused into a unified model. The Flash-Thinking-2601 upgrade added multi-environment robustness training, exposing the model to thousands of diverse environments during RL to reduce tool-use brittleness. The publisher reports that this approach significantly reduced the training cost to adapt the model to new tools compared to fine-tuning from scratch.

Key Benchmarks

Publisher-reported scores. Not independently verified by this site. Descriptions summarize what each benchmark measures.

BenchmarkScoreWhat It Measures
AIME-25American Invitational Mathematics Examination 2025 — 15-question, 3-hour competition requiring advanced mathematical reasoning
IMO-AnswerBenchInternational Mathematical Olympiad answer benchmark — tests formal mathematical reasoning on olympiad-level problems
BrowseCompComplex web browsing competition — measures ability to navigate, search, and synthesize information across multiple web pages
τ²-BenchAgentic tool-use benchmark across telecom and retail domains — measures real-world tool calling ability
LCB (LiveCodeBench)Live coding benchmark with problems released after the model's training cutoff — measures genuine coding ability, not memorization

Deployment

LongCat-Flash-Thinking is available via the LongCat API platform with a 256K maximum output limit and 128K context window. The 'Deep Thinking' function (re-thinking mode) is available for free on longcat.ai — the model activates when the deep thinking toggle is selected. Model weights are on HuggingFace (meituan-longcat/LongCat-Flash-Thinking) and GitHub (meituan-longcat/LongCat-Flash-Thinking) under the MIT License. Deployment follows the same SGLang/vLLM configuration as Flash-Chat: minimum one 8-GPU node (8x H20) for FP8, or two 8-GPU nodes (16x H800) for BF16.

Capabilities

Described in the cited primary documentation. Not an independent performance evaluation.

  • 560B MoE reasoning model with long Chain-of-Thought.
  • Domain-parallel training across STEM, Code, Agentic domains.
  • DORA asynchronous RL: >3x speedup over synchronous methods.
  • AIME-25 score 100.0 (Heavy Thinking mode); IMO-AnswerBench 86.8.
  • 64.5% fewer tokens on AIME-25 without accuracy loss.
  • Flash-Thinking-2601 upgrade adds Heavy Thinking and multi-environment robustness training.

Access & License

MIT License.

Documented context window: 128,000 tokens (per the cited publisher API documentation).

Sources

FAQ

Where can the primary source for LongCat-Flash-Thinking be found?
The official GitHub repository (meituan-longcat) and HuggingFace page contain the primary documentation, model weights, and license for LongCat-Flash-Thinking.
Is LongCat-Flash-Thinking available through the LongCat API?
LongCat-Flash-Thinking 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 2026-07-27. 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: 2026-07-27