美团发布LongCat-2.0 训练、推理全程“国产化” 性能接近Claude Opus 4.6

Meituan has launched LongCat-2.0, the industry's first trillion-parameter large language model (LLM) entirely trained and deployed on Chinese-developed computing power. It ranks among the top-tier domestically developed LLMs in China, with the following key highlights for general users and developers:

1. Core specifications and performance: Built on a Mixture-of-Experts (MoE) architecture, the model has a total of 1.6 trillion parameters, with approximately 480 billion parameters activated per token. It natively supports million-token ultra-long context inputs, is compatible with multiple mainstream development tools, and delivers particularly strong performance in coding tasks.

2. Market performance: After its preview version was anonymously launched on OpenRouter, a global LLM API routing platform, its total call volume ranked among the top three globally by the end of June. It also ranks among the top globally for call volume across multiple agent scenarios, making it one of the most popular free LLMs among developers worldwide, with performance close to that of Claude Opus 4.6.

3. Upcoming benefits: Meituan announced it will open-source core technologies including its infrastructure framework, inference engine, and model parameters in the near future, available for free access and use by developers globally.

Meituan's launch of this top-tier fully domestically developed LLM provides valuable insights for brands to track industry trends and plan business layouts:

1. Cost advantages in technology: The training and inference costs of LongCat-2.0 are lower than other comparable trillion-parameter LLMs globally. After its core technologies are open-sourced, brands will be able to develop their own intelligent marketing, customer service, and content generation tools based on the model at low cost, reducing technology investment for digital transformation.

2. New business opportunities on the platform: As a leading local life service platform in China, Meituan will inevitably roll out more intelligent merchant operation tools after rolling out its self-developed LLM to improve user matching efficiency. Brands can monitor upcoming opening plans in advance to capture new operational growth opportunities.

3. Reduced supply chain risk: The successful deployment of this fully domestically developed LLM verifies the maturity of China's domestic AI supply chain. Brands can adopt domestic solutions for their digital layouts to reduce risks stemming from changes in overseas technology policies.

The launch of Meituan's LongCat-2.0 brings new technology opportunities and risk alerts for all online and offline sellers, with key takeaways as follows:

1. Low-cost technology opportunities: The model's training and inference costs are lower than comparable LLMs globally, and all core technologies will be open-sourced. Sellers can build intelligent customer service, product copy generation, user demand analysis and other operation tools based on the model at low cost, cutting daily operational technology costs.

2. Platform growth opportunities: As a core local life trading platform in China, Meituan will upgrade its traffic distribution and merchant operation systems after launching its self-developed LLM. Sellers can follow related updates in advance, adapt to new rules early, and capture a new wave of traffic growth dividends.

3. Risk mitigation: The maturity of fully domestically developed LLMs means sellers no longer need to rely on overseas technology services to deploy intelligent tools. This eliminates the risk of service disruptions caused by changes in overseas technology policies, delivering stronger security guarantees.

The launch of Meituan's LongCat-2.0 offers multiple insights for factories advancing digital transformation and exploring business opportunities, with key highlights below:

1. Feasibility of localized digital transformation: LongCat-2.0 was trained entirely on domestically produced computing power, peaking at over 50,000 domestic AI accelerators and solving core challenges including ten-thousand-card fault tolerance and computing efficiency improvement. It verifies that domestic computing power can support large-scale intelligent R&D, allowing factories to adopt domestic solutions for digital transformation with lower costs and smaller risks.

2. Opportunities for production and design upgrades: The model supports million-token ultra-long context inputs, and stands out for its coding and computing power optimization capabilities. After its core technologies are open-sourced, factories can develop intelligent tools for industrial drawing analysis, production process optimization, and product demand processing based on the model to boost production and design efficiency.

3. Reference for cooperation models: The joint "model-chip collaborative R&D" model adopted by Meituan and domestic computing power providers provides a replicable sample for factories and technology partners to advance intelligent transformation. Factories can leverage this deep collaboration model to solve transformation pain points.

The launch of LongCat-2.0 clarifies industry development direction and brings new business opportunities for AI-related service providers, with key takeaways as follows:

1. Clear industry development trend: Domestic LLMs have now achieved full-process localized R&D for trillion-parameter models, with performance close to the global top tier and lower costs. Localization substitution has become a clear development direction for the LLM industry. Service providers can layout domestic LLM implementation services in advance to capture first-mover advantage in the segment.

2. Aligning with core client pain points: Many enterprise clients have LLM application demands, but face high costs and major risks from relying on overseas technology. LongCat-2.0 solves these exact pain points. Service providers can build customized private LLM implementation solutions for clients based on the upcoming open-source core technologies to expand new business lines.

3. Reference for technology R&D: The model introduces multiple original innovations including the industry's first zero-compute expert mechanism, which enables dynamic computing power allocation. Service providers can leverage this technology to improve the performance of their own LLM services and reduce service costs.

The launch of Meituan's LongCat-2.0 provides valuable references for all AI platforms and technology platforms in operation and ecosystem building, with key highlights below:

1. Clear direction of user demand: Performance data from the preview version shows that high-performance, cost-effective domestic LLMs are widely popular among global developers. The preview version entered the top three globally in total call volume within just two months of launch, which proves adding such models can drive large volumes of developer traffic to platforms. AI platforms can expand their supply of domestic LLMs to attract more users.

2. Reference for R&D models: The joint model-chip collaborative R&D model adopted by Meituan and domestic computing power providers solved core challenges for large-scale LLM training step by step, verifying the feasibility of collaborative R&D across China's domestic industrial chain. Platforms planning large-scale LLM R&D can adopt this model to reduce R&D risks.

3. Opportunities for ecosystem building: LongCat-2.0 will open-source its core technologies in the near future. Major AI development platforms can connect to related resources in advance, integrate the model to attract more developers, and develop supporting tool services to improve platform ecosystems and unlock new growth points.

The launch of Meituan's LongCat-2.0 provides a new research sample for LLM industry research, with multiple notable new trends worth studying. Key highlights are as follows:

1. New industry trends: This is the first trillion-parameter LLM in the industry to complete the full training and inference process entirely on domestic computing power, representing the largest-scale training task ever completed on domestic computing infrastructure. It solved core challenges including ten-thousand-card fault tolerance and improved computing utilization, marking that China's full domestic LLM industrial chain already has the capacity to support trillion-parameter LLM R&D, with industry growth outpacing expectations.

2. Technical innovation research: The model introduced the industry's first zero-compute expert mechanism, enabling token-level dynamic computing allocation that effectively reduces training and inference costs. It also includes multiple other original designs such as a cross-layer fast connection architecture, opening up new research directions for LLM computing power optimization.

3. Model innovation research: Meituan's "model-chip collaborative" joint R&D model, alongside its approach of open-sourcing core technologies to give back to the developer community, provides a new research sample for industry-university-research translation in LLM development, with high research value.