LongCat‑2.0: Training a Trillion‑Parameter Model on a Domestic 50k‑Card Cluster

On June 30, Meituan released LongCat‑2.0, a 1.6‑trillion‑parameter mixture‑of‑experts (MoE) model that was trained and inferred end‑to‑end on a domestic 50,000‑card compute cluster and is now open‑sourced.

Training Scale and Stability

The pre‑training dataset exceeds 30 TB of tokens covering Chinese, English, multilingual text and code. To cope with hardware failures, communication anomalies, memory pressure and numerical instability at the ten‑thousand‑card scale, the team improved three aspects:

Stability: handling inter‑card communication errors, elastic scaling and automatic fault recovery reduced the monthly fault rate by more than 70%.

Correctness: custom deterministic operators, bitwise consistency checks and parameter validation ensured reliable training results while improving key module precision and Reduce logic.

Efficiency: pipeline scheduling, memory optimizations and operator‑level core binding raised the training MFU by 1.5×.

Consequently, LongCat‑2.0 sustains a steady throughput of over 1 T tokens per day, confirming stable trillion‑parameter training on domestic hardware.

Inference Optimizations

During inference, model, operator and framework co‑optimizations were applied:

Large‑scale expert parallelism aggregates memory bandwidth to support low‑latency decoding of the MoE model.

Zero‑compute expert routing skips computation for simple tokens, avoiding unnecessary transmission and calculation.

Communication, Attention and GEMM operators received dedicated scheduling tweaks, and framework mechanisms such as early weight prefetch further reduced waiting overhead.

These measures enable the model to run stably in real‑world agentic coding tasks.

Architecture for Agentic Coding

The design centers on making the model more efficient and stable for real agentic coding workloads.

1 M‑token context: LongCat Sparse Attention (LSA) replaces quadratic attention with linear‑complexity sparse attention, allowing the model to retain precise information over up to one million tokens.

Zero‑compute expert + ScMoE: Token‑level dynamic activation (33 B–56 B active parameters) allocates compute only to complex tokens, keeping simple tokens cheap.

MOPD multi‑expert fusion: Three expert groups—Agent, Reasoning, Interaction—are gated at inference time to the most suitable expert set, rather than naïvely merging parameters. This yields strong performance across coding, reasoning and interactive tasks.

Benchmark Results

Comprehensive evaluations show LongCat‑2.0 excels in both coding ability and complex office‑task handling.

On SWE‑bench Pro, it scores 59.5, surpassing Gemini 3.1 Pro (54.2), GPT‑5.5 (58.6) and Claude Opus 4.6 (57.3).

On SWE‑bench Multilingual, it achieves 77.3, comparable to Claude Opus 4.6 (77.8).

In Terminal‑Bench 2.1 (real‑world terminal interaction), it reaches 70.8, demonstrating stable execution and error correction.

For real‑office agent tasks, it attains 78.8 on RWSearch, 73.2 on FORTE and 79.9 on BrowseComp, matching or approaching leading closed‑source models.

Real‑World Applications

During internal testing, LongCat‑2.0 was used to build AI agents for various scenarios:

AI SQL Agent: natural‑language queries are understood, planned, executed and returned as business insights.

Codebase migration: given legacy plugin code and a new SDK, the model rewrites the entire plugin to the new API while preserving functionality.

Full‑app generation: from a single prompt describing a children’s AI game, the model produces architecture, logic and visual details for three playable pages.

3D interactive demo: a one‑sentence description yields a complete Three.js scene with interactive fluid and particle effects.

AI novel factory: the model orchestrates multiple agents to generate world‑building, chapter writing, quality assessment and revision, maintaining consistency over million‑word narratives.

Open Access and Adoption

The preview version is already available via the OpenRouter platform and longcat.ai, ranking among the top three global model call volumes (behind Hermes and Claude Code). The API is publicly listed at https://longcat.chat/platform/product.

Overall, LongCat‑2.0 demonstrates that Meituan can not only train a trillion‑parameter model on domestic hardware but also deliver a stable, high‑performance system for real‑world agentic coding tasks.