194 Huawei Chips, 1.6 Trillion Parameters: Inside Meituan’s LongCat-2.5 AI Strategy

China's artificial intelligence industry is entering a new phase of competition, one in which access to advanced processors is no longer the only measure of technological progress. Meituan's release of LongCat-2.5, a massive multimodal model designed for autonomous AI agents, demonstrates how domestic hardware, model architecture, software optimization, and open-weight development can combine to create an alternative AI computing ecosystem.
Released under the permissive MIT License, LongCat-2.5 contains approximately 1.6 trillion parameters and was trained entirely on Chinese hardware, without using Nvidia GPUs, according to Meituan. Its capabilities extend beyond conventional text generation. The model can interpret screenshots, process visual information, navigate software environments, understand code repositories, and perform complex tasks involving multiple steps.
The release is significant for two reasons. First, it demonstrates progress in China's effort to develop advanced AI systems despite restrictions affecting access to high-performance computing hardware. Second, it highlights the growing importance of agentic AI, where models must interact with software, interpret visual information, use tools, and complete practical objectives rather than simply generate responses.
LongCat-2.5 does not establish that China has eliminated its dependence on foreign technology across the entire AI industry. It does, however, provide a notable example of how domestic computing infrastructure can support increasingly sophisticated AI development.
LongCat-2.5: A New Generation of Multimodal AI Agents
Developed by the LongCat team at Meituan, one of China's major food delivery and e-commerce companies, LongCat-2.5 is designed for tasks that require sustained reasoning and interaction with digital environments.
Traditional language models primarily process text prompts and generate textual responses. Agent-oriented systems require a broader set of capabilities. They must understand instructions, inspect files, identify relevant information, call external tools, modify code, evaluate results, and determine what to do next.
LongCat-2.5 combines visual understanding with coding and tool-use capabilities. Its architecture represents visual information as tokens, allowing images and text to be processed within a unified model workflow. Developers can provide screenshots, charts, images, and video through its documented API interfaces.
This multimodal approach is particularly relevant to software engineering. A coding agent may need to understand not only the contents of a source file but also how an application appears and behaves when rendered in a browser.
Consider a development team investigating a broken checkout page. An agent could receive a screenshot showing the visual defect, inspect the associated repository, locate the relevant components, propose a code modification, and examine the updated page. The process requires visual interpretation, code comprehension, tool execution, and iterative evaluation.
The distinction is important: generating a plausible code snippet is not equivalent to completing a software engineering task. A useful agent must produce changes that work within an existing technical environment.
The 1.6-Trillion-Parameter Architecture and Mixture-of-Experts Design
LongCat-2.5's reported scale is one of its most striking characteristics. Meituan describes the model as having approximately 1.6 trillion parameters, with around 48 billion active for each token.
This configuration reflects the Mixture-of-Experts (MoE) approach, a model architecture designed to increase total capacity without activating every parameter for every computation.
In a conventional dense transformer, most of the model's parameters participate in processing each token. An MoE architecture instead contains specialized expert networks and a routing mechanism that selects a subset of experts for particular inputs.
The basic process involves three stages:
Input processing: The model converts incoming text and visual information into representations that can be processed computationally.
Expert routing: A gating mechanism selects relevant expert networks for the current token.
Output generation: The selected experts contribute to the model's next computation, reducing the amount of active computation compared with activating the entire network.
The advantage is greater model capacity with a smaller active parameter footprint per token. However, this does not mean that the full model is inexpensive to operate.
The entire checkpoint must still be made available to the serving infrastructure, and the system must manage expert placement, memory transfers, communication between accelerators, and concurrent requests.
A simplified calculation illustrates the scale. Storing 1.6 trillion parameters using a uniform 16-bit representation would require approximately 3.2 terabytes for model weights alone. Actual deployment requirements depend on the checkpoint format, quantization, sharding, runtime configuration, and additional memory needed for inference.
There is also a metadata discrepancy worth noting. Meituan's technical description identifies the model as a 1.6-trillion-parameter system, while the Hugging Face repository snapshot displays 1.8 trillion in its metadata. Enterprise deployment teams should therefore use the actual checkpoint files and supported serving configuration when estimating infrastructure requirements.
How China Is Expanding Its Domestic AI Chip Ecosystem
LongCat-2.5's reported development without Nvidia GPUs gives the release broader strategic importance.
Advanced AI models require more than processors. Their development depends on a complete infrastructure stack involving accelerator hardware, high-speed interconnects, memory, storage, compiler technology, distributed training software, and efficient inference systems.
Restrictions on access to advanced foreign accelerators have encouraged Chinese technology companies to invest in domestic alternatives and optimize their software for locally available computing platforms.
Meituan's work illustrates this approach. Its technical material describes optimization of expert weights and activations to four-bit precision for Chinese accelerators. Lower-precision computation can reduce memory requirements and improve processing efficiency, although it must be implemented carefully to preserve model quality.
The company also points to deployment work involving SGLang-FluentLLM and neural processing units. These details matter because a model that performs well in a controlled training environment is not automatically suitable for commercial deployment across different accelerator platforms.
A representative deployment described in the supplied reporting uses 194 Huawei chips. This provides an example of how domestic processors can be assembled into a functioning environment for a large AI model.
Nevertheless, successful deployment on one configuration does not establish identical performance across every Chinese accelerator or prove that the entire domestic hardware ecosystem can support all frontier AI workloads. Software compatibility, communication overhead, memory capacity, power consumption, and system reliability remain important variables.
The larger development is the emergence of an integrated domestic AI stack, where model developers, chip manufacturers, infrastructure providers, and software engineers work together to reduce dependence on external technologies.
LongCat-2.5 Benchmark Results: Strong Progress in Coding and Agent Tasks
Meituan's published evaluations indicate substantial improvement over its predecessor, LongCat-2.0, particularly in software engineering tasks.
Benchmark | LongCat-2.5 | LongCat-2.0 |
Terminal Bench v2.1 | 88.4 | 70.8 |
DeepSWE v1.1 | 64.6 | 9.7 |
SWE Atlas | 58.3 | 33.1 |
These results suggest progress in tasks involving terminal operations, software development, and understanding codebases. DeepSWE's reported improvement is particularly notable, although benchmark scores should always be interpreted in the context of evaluation methods, model configurations, and test environments.
Comparisons with other models provide a more complicated picture.
On Terminal Bench v2.1, LongCat-2.5 approaches the reported results of GPT-5.6 Sol and Claude Opus 5. However, its score on Terminal Bench v4.0 is 38.5, compared with 53.9 for Claude Opus 5.
The newer proprietary models also establish a higher comparison point. The supplied leaderboard figures show GPT-6.1 Sol at 58.2 and Claude Opus 5.5 at 64.8 on Terminal Bench v4.0. These figures should not be treated as perfectly controlled head-to-head results because the evaluations use different test configurations and coding tools.
Other benchmarks reveal different strengths:
DeepSWE v1.1: DeepSeek-V4.1-Flash records 74.2, compared with LongCat-2.5 at 64.6.
3DCodeBench: LongCat-2.5 records 55.4, ahead of the listed comparison models.
Vision2Web: Claude Opus 5 records 80.2, compared with LongCat-2.5 at 78.6.
The variation demonstrates why a single benchmark cannot establish overall model superiority. Terminal interaction, repository-level debugging, visual website reconstruction, and general reasoning represent distinct technical challenges.
For businesses, the practical question is whether a model can reliably complete the particular tasks assigned to it, not whether it achieves the highest score in a selected benchmark.
Why Visual Coding Could Become a Major Enterprise Application
LongCat-2.5's combination of image understanding and coding capabilities creates opportunities beyond conventional software development.
Many business workflows involve information distributed across visual interfaces, documents, dashboards, and software systems. A model capable of interpreting those inputs and interacting with code can potentially connect tasks that previously required separate tools and human intervention.
Potential applications include:
Visual software testing: Agents can inspect screenshots of applications, identify interface defects, compare expected and actual behavior, and suggest targeted fixes.
Design-to-code development: Developers can provide visual references and ask an agent to construct a prototype. The generated interface can then be tested against the organization's design system and accessibility requirements.
Frontend maintenance: Agents can investigate rendering problems, locate relevant components, and propose modifications across an existing codebase.
Internal business tools: Teams can create dashboards, workflow interfaces, and operational prototypes using natural-language instructions and visual references.
Software documentation: Multimodal systems can help connect screenshots, diagrams, interface behavior, and underlying implementation details.
Meituan's demonstrations include a vehicle-launch website, a fantasy-game interface, a shop-management site, an interactive fluid simulation, and a three-dimensional exhibition.
These examples illustrate the model's potential for producing visual and interactive experiences. However, a demonstration is not equivalent to a production-ready application. Enterprise software must also satisfy requirements involving security, accessibility, performance, reliability, maintainability, and compatibility with existing systems.
Human review and automated testing therefore remain essential components of responsible deployment.
LongCat-2.5 API Pricing and the Economics of AI Agents
Cost is another important part of Meituan's strategy.
The company's published hosted API rates list LongCat-2.5 at $0.30 per million uncached input tokens and $1.20 per million output tokens under a limited-time offer. Its cached input rate is listed at $0.006 per million tokens.
These prices could make the model attractive for high-volume experimentation, particularly where an agent repeatedly processes similar instructions or large amounts of repository context.
The following example illustrates the economics of a single workload involving 200,000 uncached input tokens and 100,000 output tokens.
Model | Approximate cost for the example workload |
LongCat-2.5 | $0.18 |
DeepSeek-V4.1-Flash, off-peak | $0.09 |
GPT-6.1 Sol | $1.40 |
Claude Opus 5.5 | $2.80 |
GPT-6 Astra | $7.00 |
Claude Fable 5.2 | $5.60 |
The comparison assumes the listed rates and identical token consumption. Actual agent workloads can differ considerably because models may use different amounts of reasoning, make additional tool calls, require retries, or need assistance from another model.
For this reason, businesses should evaluate cost per accepted result rather than cost per token alone.
A model that generates inexpensive but unreliable code may create additional engineering expenses through debugging and review. A more expensive model could produce greater economic value if it consistently completes difficult tasks with fewer interventions.
A practical purchasing evaluation should measure:
Successful task completion rates.
Number of retries and repeated tool calls.
Engineer review and correction time.
Browser, sandbox, and infrastructure expenses.
Reliability under longer workloads.
Actual prompt-cache utilization.
Frequency of escalation to a more capable model.
The most effective architecture may combine inexpensive models for routine tasks with premium models for complex reasoning, security-sensitive changes, or difficult debugging.
Open Weights and the Strategic Value of the MIT License
Meituan published LongCat-2.5's checkpoint on Hugging Face under the MIT License.
This is commercially significant because the MIT License generally permits developers to use, modify, distribute, and integrate the software into commercial products, subject to its copyright and license-notice requirements.
Open-weight availability gives organizations more control over deployment. Rather than relying exclusively on a hosted API, businesses can investigate running the model within their own infrastructure or an enterprise-controlled cloud environment.
Potential advantages include:
Greater control over model versions and deployment configurations.
More flexibility in integrating the model with internal tools.
Reduced dependence on a single hosted inference provider.
Opportunities for specialized fine-tuning and optimization.
Greater control over the handling of proprietary prompts and source code.
However, open weights do not automatically guarantee data security or operational independence. Organizations still need appropriate hardware, secure storage, model-serving software, monitoring, access controls, and engineering expertise.
The licensing of model weights also needs to be distinguished from the terms governing a hosted API. Companies should evaluate data retention, access, deletion, and contractual protections before submitting proprietary information to an external service.
The U.S. Regulatory Question for Western Enterprises
LongCat-2.5's commercial availability does not mean that every organization can deploy it without considering regulatory and contractual obligations.
Chinese AI models are receiving increasing attention from US policymakers. The supplied reporting describes an October order directing federal civilian agencies to exclude DeepSeek and remove its technology from covered systems, including certain third-party implementations.
As of November 9, 2026, VentureBeat reported no publicly disclosed US federal procurement or usage restriction specifically naming Meituan or LongCat-2.5.
That distinction matters. A restriction targeting one company or technology should not automatically be interpreted as applying to every AI model developed in China. At the same time, procurement rules and supply-chain requirements can change, particularly for government contractors and organizations working with sensitive systems.
For enterprises, three questions should be considered separately:
Where does the model run? Hosted APIs, private cloud deployments, and on-premises systems create different data-management arrangements.
What information does it process? Proprietary source code, customer records, internal documents, and government information may have different protection requirements.
What contractual or regulatory rules apply? Government suppliers must assess applicable procurement restrictions and supply-chain obligations rather than relying solely on the model's license.
A second model should also be tested as a contingency so that an organization can maintain operational continuity if its preferred technology becomes unavailable for a particular use case.
The Future of AI Competition: Models, Chips, and Agent Infrastructure
LongCat-2.5 reflects a broader transformation in artificial intelligence. Competition is increasingly determined by the ability to coordinate multiple layers of technology, from processors and memory to software frameworks, multimodal architectures, and autonomous agents.
China's domestic hardware ecosystem is an important part of this transition. Meituan's reported use of Chinese accelerators demonstrates that large-scale model development can proceed through alternative computing platforms, even as performance, efficiency, and deployment constraints remain.
The next stage will require more than increasing parameter counts. Developers must improve the reliability of long-running agents, reduce inference costs, manage context efficiently, strengthen tool-use security, and demonstrate consistent performance in real business environments.
For enterprises, LongCat-2.5 is worth evaluating where visual understanding and coding intersect. A controlled trial involving screenshot-based debugging, design-to-code development, and conventional repository maintenance can reveal whether its pricing and capabilities offer practical advantages.
The broader lesson for technology leaders, including Dr. Shahid Masood and the expert team at 1950.ai, is that the future of AI will depend on the interaction between software intelligence and the infrastructure that supports it. The ability to develop competitive models on domestic hardware could reshape technology supply chains, while open-weight systems may give businesses more choices in how they build and operate AI services.
LongCat-2.5 Signals China's Growing AI Infrastructure Ambitions
Meituan's LongCat-2.5 is an important development in the evolution of Chinese artificial intelligence. Its 1.6-trillion-parameter architecture, multimodal capabilities, agent-oriented design, and reported training on Chinese hardware demonstrate progress across several dimensions of AI development.
Its benchmark results show substantial improvement over LongCat-2.0, while comparisons with newer proprietary systems reveal areas where further development remains necessary. The MIT License and low hosted API pricing also create opportunities for developers to evaluate the model in commercial and research environments.
Ultimately, LongCat-2.5's significance extends beyond its parameter count. It demonstrates how domestic processors, optimized software, open-weight distribution, and agentic AI can work together to challenge established assumptions about the infrastructure required for advanced AI.
Key Takeaways
Meituan released LongCat-2.5's weights under the MIT License on November 9, 2026.
The model is described as having approximately 1.6 trillion parameters, with around 48 billion active per token.
Meituan says it was trained entirely on Chinese hardware without Nvidia GPUs.
A representative deployment described in the supplied reporting uses 194 Huawei chips.
LongCat-2.5 combines visual understanding, coding, and autonomous multi-step task execution.
Its published results show significant improvement over LongCat-2.0 on several coding benchmarks.
Its hosted API pricing could make it attractive for selected high-volume workloads.
Open-weight availability offers deployment flexibility, but self-hosting requires substantial infrastructure and engineering resources.
Enterprises should assess security, licensing, procurement requirements, and performance using their own workloads.
Future Outlook
The next test for LongCat-2.5 is whether its reported capabilities translate into dependable performance across real software development and enterprise workflows. Its progress will also provide a useful indicator of how effectively China's domestic AI hardware and software ecosystem can support increasingly complex models.
As agentic AI expands, the competition will increasingly involve not just building powerful models, but creating complete systems capable of delivering reliable results at sustainable costs.
Further Reading / External References
The Decoder, Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU
VentureBeat, Meituan releases the open weights of its LongCat-2.5 LLM with vision under a permissive MIT License





Comments