LongCat-Flash-Thinking-2601 Technical Report

TL;DR

Meituan’s LongCat team releases LongCat-Flash-Thinking-2601, a 560B-total / 27B-activated open-source MoE reasoning model purpose-built for agentic reasoning (agentic search, tool use, tool-integrated reasoning). The recipe consolidates three things this team’s wiki has tracked separately: (a) a unified post-training pipeline that trains domain-specialized expert models under a shared framework and then fuses them at both the model-level and the data-level into one general model, (b) a “Heavy Thinking” test-time scaling mode that expands reasoning width (parallel trajectories) and depth (iterative summarization) jointly, (c) an “environment scaling” pipeline that auto-builds 10,000+ executable, verifiable environments across 20+ domains as the substrate for RL. Also ships a “Zigzag Attention” variant (MLA + SSA mix) introduced mid-training for 1.5× speedup at 1M-token contexts. Open weights on Hugging Face; code on GitHub.

Key claims

LongCat-Flash-Thinking-2601 is a 560B-total / 27B-activated-per-token MoE reasoning model claiming SOTA among open-source models across agentic search, agentic tool use, tool-integrated reasoning, mathematical reasoning, and coding [Abstract, §1].
The post-training pipeline is a domain-parallel experts + fusion recipe: domain-specialized expert models are first trained under a shared framework, then consolidated into a single general model through both model-level and data-level merging [§1, “unified multi-domain post-training pipeline”].
Heavy Thinking Mode is a test-time scaling mechanism that jointly expands reasoning width (parallel trajectory exploration) and depth (iterative reasoning refinement via a summary model that reflects on and synthesizes parallel trajectories), plus a purpose-built context-memory module to keep reasoning coherent over long horizons [§1, Heavy Thinking framework].
Environment scaling for agentic RL: an automated pipeline that builds 10,000+ executable, verifiable environments across 20+ domains as the substrate for tool-use RL [Abstract; release blog]. The reported design bet is that “in-depth exploration of environment scaling and principled task construction” drives the model’s generalization in complex tool use [Abstract].
RL infrastructure: an upgraded DORA framework supports async training with 32,000+ concurrent environments, targeted at stability issues in long-tail and highly heterogeneous tasks [release blog, “RL infrastructure”].
Noise injection and curriculum RL address “greenhouse-agent” overfitting — real-world user/tool noise is systematically analyzed and injected into the training loop; a curriculum-based strategy gradually toughens the model against messy, imperfect environments [release blog, “Robustness in the wild”].
A Zigzag Attention architectural variant — Zigzag Connectivity combining MLA (Multi-head Latent Attention) and SSA (Sparse / Sliding Sparse Attention) — is introduced via a mid-training switch to sparse variants, reportedly yielding a 1.5× speedup and supporting 1M-token contexts [release blog, “Zigzag Attention”].
Pre-training inherits the LongCat-Flash-Chat (Meituan et al., 2025) recipe to preserve general-purpose language and reasoning capabilities; mid-training uses a staged procedure with progressively increasing context lengths to support long-horizon agentic trajectories [§1].
Evaluations include Olympiad-level math benchmarks AIME 2025, HMMT 2025 (February), IMO-AnswerBench, and a new AMO-Bench [§Evaluation; per HTML §math].

Method

The model is a 560B-total / 27B-activated MoE. Pre-training continues from LongCat-Flash-Chat’s distribution to retain general capability. A staged mid-training phase progressively expands context length and switches from dense attention variants to Zigzag Attention — a connectivity pattern that interleaves MLA and SSA blocks — reportedly yielding a ~1.5× speedup and unlocking 1M-token contexts.

Post-training is a unified multi-domain pipeline: separate domain-specialized expert models (math, coding, agentic search, tool use, …) are trained under a shared framework, then merged into a single general model through both model-level parameter merging and data-level corpus consolidation. For agentic capabilities specifically, the team automates construction of >10,000 executable, verifiable RL environments across >20 domains, and runs large-scale GRPO-family RL on the upgraded DORA infrastructure (asynchronous, with 32,000+ concurrent environments). Robustness is targeted by noise injection (user/tool noise sampled from real interaction logs, injected directly into the training loop) and curriculum RL (progressively harder, messier environments).

At inference, the model exposes a Heavy Thinking mode for test-time scaling: parallel trajectory exploration produces multiple independent reasoning chains, an iterative-summarization model reflects on and synthesizes them, and a dedicated context-memory module preserves coherence across the resulting wide+deep reasoning graph.

Results

Claims SOTA among open-source models on a wide range of agentic benchmarks — agentic search, agentic tool use, tool-integrated reasoning [Abstract]. Specific benchmark numbers and head-to-head tables are reported in the released technical report and benchmark figure (Fig. 1) but not extracted here.
Math: evaluations on AIME 2025, HMMT 2025 (February), IMO-AnswerBench, and AMO-Bench [§math benchmarks]. Headline numbers are in the report’s results tables (not extracted in fetched material).
The Zigzag-Attention sparse variant is reported to give ~1.5× inference speedup at 1M-token contexts vs. the dense baseline [release blog].
14 pages of “deep dives” on large-scale agentic RL are explicitly highlighted in the release blog as the most novel contribution.

Why it’s interesting

This is the largest open-weights reasoning MoE filed so far in this wiki (560B-total / 27B-activated vs. HunyuanImage 3.0’s 80B-total / 13B-activated in image generation; Kimi K2.5 is a closer agentic counterpart but doesn’t expose the same training recipe in this much detail). It bundles together themes the wiki tracks separately: the reasoning-RL consensus that GRPO-family RL on shaped rewards over self-synthesized trajectories is the post-training default (Reasoning RL); the tool-use-agents (Tool-Use Agents) shift toward synthesized agentic environments at scale; and a test-time-scaling (Inference-Time Scaling) mode (Heavy Thinking) that combines parallel width and iterative summarization — sibling to Kimi K2.5’s PARL and to RLM-style scaffold scaling.

The two most distinctive contributions look like (a) the domain-parallel experts + model-and-data-level fusion post-training pattern, which is a more explicit architectural commitment than MiroThinker’s three-stage SFT→DPO→GRPO; and (b) environment scaling as a first-class lever — auto-building 10,000+ verifiable RL environments across 20+ domains, with the DORA infra supporting 32K concurrent envs. This is a sharper version of InfTool’s multi-agent role-playing loop (Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing) — instead of synthesizing trajectories in a fixed environment, the synthesis target is the environment itself. Noise injection + curriculum RL targeting “greenhouse-agent” overfitting is a directly testable robustness recipe that the wiki’s existing agentic-RL entries have not isolated.