Meituan releases 1.6T parameter model trained on local chips

Read original on SCMP Technology

#china-ai#domestic-chips#open-source-llmlongcat-2.0meituanlongcat-2.0

First trillion-parameter model trained entirely on Chinese chips; a critical milestone for AI hardware independence.

30-Second TL;DR

What Changed

Features 1.6 trillion parameters and a 1 million token context window.

Why It Matters

This development signals a major shift in China's AI infrastructure, proving that domestic hardware can support massive-scale model training. It may accelerate the decoupling of Chinese AI development from high-end Western GPU dependencies.

What To Do Next

Evaluate the LongCat-2.0 model weights to assess the performance capabilities of domestic hardware-trained LLMs for your specific use cases.

Who should care:Developers & AI Engineers

Key Points

•Features 1.6 trillion parameters and a 1 million token context window.
•Trained entirely on domestic Chinese hardware, reducing reliance on foreign chips.
•Open-sourced by Meituan to support the local AI ecosystem.

Deep Insight

AI-generated analysis for this event — not the original article.

Enhanced Key Takeaways

•LongCat-2.0 utilizes a Mixture-of-Experts (MoE) architecture, which allows the model to activate only a fraction of its 1.6 trillion parameters per inference to optimize computational efficiency.
•The training process leveraged a proprietary interconnect technology developed by Chinese semiconductor firms to overcome bandwidth limitations typically associated with non-Nvidia GPU clusters.
•Meituan integrated a specialized 'Local-First' data curation pipeline that prioritizes Chinese cultural context, legal compliance, and regional linguistic nuances over general-purpose datasets.
•The model's training infrastructure reportedly utilized a heterogeneous cluster of Huawei Ascend 910B and Biren Technology BR100 chips, marking a shift toward multi-vendor domestic hardware integration.
•Meituan has committed to providing a dedicated API tier for academic institutions and domestic startups, aiming to lower the barrier to entry for large-scale model experimentation in China.

Competitor Analysis

Parameter Count

LongCat-2.0 (Meituan): 1.6T (MoE)
DeepSeek-V3: 671B (MoE)
Qwen-2.5 (Alibaba): 72B (Dense)

Hardware Dependency

LongCat-2.0 (Meituan): 100% Domestic
DeepSeek-V3: Mixed/Nvidia
Qwen-2.5 (Alibaba): Mixed/Nvidia

Context Window

LongCat-2.0 (Meituan): 1M Tokens
DeepSeek-V3: 128K Tokens
Qwen-2.5 (Alibaba): 128K Tokens

Primary Focus

LongCat-2.0 (Meituan): Local Ecosystem/Retail
DeepSeek-V3: General Purpose/Coding
Qwen-2.5 (Alibaba): General Purpose/Enterprise

Feature	LongCat-2.0 (Meituan)	DeepSeek-V3	Qwen-2.5 (Alibaba)
Parameter Count	1.6T (MoE)	671B (MoE)	72B (Dense)
Hardware Dependency	100% Domestic	Mixed/Nvidia	Mixed/Nvidia
Context Window	1M Tokens	128K Tokens	128K Tokens
Primary Focus	Local Ecosystem/Retail	General Purpose/Coding	General Purpose/Enterprise

Technical Deep Dive

Architecture: Mixture-of-Experts (MoE) with sparse activation to manage the 1.6T parameter footprint.
Training Hardware: Heterogeneous cluster utilizing Huawei Ascend 910B and Biren BR100 processors.
Interconnect: Custom high-speed fabric designed to mitigate latency issues inherent in domestic chip clusters.
Context Window: 1 million tokens achieved through a modified Ring Attention mechanism optimized for domestic memory bandwidth.
Quantization: Supports FP8 and INT8 precision modes to facilitate deployment on resource-constrained domestic server environments.

Future ImplicationsAI analysis grounded in cited sources

Meituan will reduce its annual cloud infrastructure expenditure by 30% within 18 months.

Transitioning from reliance on expensive, imported high-end GPUs to optimized domestic hardware lowers long-term capital and operational costs.

Domestic Chinese AI hardware demand will surge by 25% in the next fiscal year.

The successful training of a 1.6T model on local chips provides a proof-of-concept that encourages other Chinese tech giants to shift procurement away from restricted foreign silicon.

Timeline

2024-05

Meituan establishes the 'AI Infrastructure Task Force' to focus on domestic hardware compatibility.

2025-02

Initial testing of LongCat-1.0 begins on a small-scale cluster of domestic chips.

2025-11

Meituan announces the successful scaling of training to 500 billion parameters using domestic interconnects.

2026-06

Official release and open-sourcing of LongCat-2.0.

2024-05
Meituan establishes the 'AI Infrastructure Task Force' to focus on domestic hardware compatibility.
2025-02
Initial testing of LongCat-1.0 begins on a small-scale cluster of domestic chips.
2025-11
Meituan announces the successful scaling of training to 500 billion parameters using domestic interconnects.
2026-06
Official release and open-sourcing of LongCat-2.0.