Meituan Trains a 1.6T-Parameter MoE Model on 50,000 Chinese Chips and Open-Sources It

1. Core Basic Information

Release Date: June 30, 2026 Positioning: Meituan's new generation trillion-parameter MoE foundational large model, focusing on code, agent, and ultra-long text scenarios, trained and inferred entirely on a domestic computing power stack, with a simultaneous announcement of open-sourcing core technologies. Pre-release buildup: A Preview version was launched at the end of April 2026 and anonymously connected to the global API platform OpenRouter. Within two months of going live, its call volume surged into the global top three.

Core Hardware Milestone (Industry-First Breakthrough)

The industry's first trillion-parameter large model to complete full pre-training and online inference on a cluster of 50,000 domestic AI accelerator cards, without using any overseas high-end GPUs throughout the entire process. It overcame challenges in 10,000-card cluster fault tolerance, NPU deterministic computation, and computing power utilization, validating that domestic computing power can support the training of ultra-large-scale frontier models.

2. Model Architecture and Specifications

1. Parameter Scale
   • Total Parameters: 1.6 trillion (1.6T) MoE Mixture-of-Experts architecture
   • Dynamic Activation per Token: Average 48 billion (dynamic range 33B~56B), with a self-developed Zero-Computation Expert Mechanism: simple text tokens are allocated less computing power, while complex code/long text automatically expands computing resources, significantly reducing inference costs.
2. Context Capability Natively supports 1M (million Token) ultra-long context, processing million-word documents, complete code repositories, and full project source code in one go, without context truncation or memory loss issues.
3. Training Data Over 30 trillion Tokens of pre-training data, covering Chinese and English multilingual content, massive open-source code, tool invocation, and agent interaction data.
4. Original Technical Architecture ScMoE cross-layer shortcut connections, Ngram Embedding enhancement, LongCat proprietary long-text attention, and MOPD multi-expert fusion mechanism, specifically optimized for code and agent scenarios.

3. Core Capability Highlights

1. Top-Tier Code/Development Capabilities

Deeply adapted to mainstream code agent frameworks like Claude Code, OpenClaw, and Hermes, excelling at:

• Whole-repository code understanding, batch refactoring, and bug fixing
• Multi-file linked generation for complex projects, full-stack project output
• Strong performance in professional code benchmarks like SWE-bench Pro; the Preview version ranked 2nd globally in monthly call volume on code agent platforms.

2. Native Agent Reinforcement

Optimized for automated tasks, multi-tool invocation, and complex long-chain planning, the Preview version ranked first globally in call volume on the Hermes agent platform, suitable for automation development, data analysis, and batch document processing robots.

3. Extremely Low API Call Costs

Official low-price strategy:

• Cache-hit input costs only 0.04 yuan per million Tokens; output is about 8 yuan per million Tokens, lower than comparable overseas flagship models, significantly lowering the barrier to entry for developers.

4. Global Market Performance (Preview Version Data)

1. Connected to the OpenRouter global large model routing platform, total call volume entered the global top three;
2. Rankings in specific scenarios:
   • Hermes General Agent: 1st globally in monthly call volume
   • Claude Code Code Agent: 2nd globally in monthly call volume
   • OpenClaw Development Framework: Among the global leaders
3. The official website longcat.ai has opened API access channels for global developers.

5. Open-Source Plan

Meituan announced it will fully open-source three core assets soon:

1. LongCat's underlying distributed training/inference Infra framework;
2. The self-developed inference engine adapted for domestic computing power;
3. The complete LongCat-2.0 model weights (the weight package is currently being uploaded); The goal is to lower the technical barrier for domestic developers to train trillion-level large models and promote the synergistic development of the domestic computing power and large model ecosystem.

6. Industry Significance

1. Computing Power Independence Breakthrough: Proves that a purely domestic 50,000-card cluster can stably complete the full training of a trillion-parameter model, breaking free from dependence on overseas high-end GPUs;
2. Vertical Scenario Benchmark: A large company's self-developed large model deeply rooted in the code and agent tracks, distinct from general conversational models, targeting developer and enterprise R&D scenarios;
3. Inclusive AI: Extremely low API pricing plus full open-sourcing reduces the cost of using and secondary development of trillion-level large models;
4. Meituan's Technical Layout: Leveraging massive local life service data and a self-developed large model, it can subsequently be implemented in internal business scenarios such as merchant operation intelligent assistants, delivery scheduling AI, local knowledge base customer service, and R&D automation tools.