Meituan's newly open-sourced LongCat-2.5 comes with two clear numbers: 1.6 trillion total parameters and roughly 48 billion active parameters per token. Our view is that competition in large language models is shifting from “who can build the bigger model” to “who can deploy a complex model in real business operations.”
This is a multimodal model designed for long-horizon agentic tasks. By “long-horizon agent,” we mean a system that can continuously read materials, call tools, and deliver a complete result rather than answer a single question. It supports a 1 million-token context window and is available under the MIT License, allowing commercial inference deployment and further training without additional licensing.
What This Is
LongCat-2.5 focuses on four areas: complex codebases and full project development; bringing visual understanding, code generation, and 3D design into one model; handling professional office tasks such as reports and presentations; and using W4A4 to quantize expert weights and activations to 4 bits, reducing inference costs for a trillion-parameter model.
Meituan also disclosed adaptations for domestic computing infrastructure. With less than 80GB of VRAM per card, it conducted long-sequence training and fine-tuning, while its FSDP configuration delivered 86% higher single-card throughput than a DDP baseline. These figures show that deployment constraints are moving upstream into model and infrastructure design.
Industry View
For Meituan, an open-source model has greater strategic value than simply launching an API: it can provide the technical foundation for cloud services and internal operations while also expanding the ecosystem through developer feedback. The permissive MIT License further lowers the legal cost for enterprises to evaluate, modify, and deploy the model.
The risks are just as clear. Meituan says downstream task performance after quantization is comparable to that before quantization, but the original announcement did not disclose complete scores, cost comparisons, or testing conditions. Official demonstrations also cannot replace validation on each enterprise's own data. Even with low-bit quantization, a trillion-parameter model still places substantial demands on memory capacity, bandwidth, and operational expertise, making self-deployment unsuitable for many small and medium-sized businesses. What truly warrants attention is whether third parties can consistently reproduce its performance on long-running tasks, as well as its real-world throughput and failure rates on domestic chips.
Impact on Regular People
For enterprise IT: Businesses gain a domestic open-source option that can be privately deployed and customized. They can also evaluate Meituan's model alongside other solutions under the same framework covering cost, permissions, and stability.
For individual professionals: Complex reports, code projects, and cross-tool collaboration may increasingly be delegated to agents. Critical conclusions will still require human review, however, and teams will need rules for call logs, data boundaries, and result acceptance.
For the consumer market: If Meituan integrates these capabilities into merchant tools, customer service, or local-life services, end users may never notice the model's name. They may simply receive automated services sooner, with less waiting and more concrete deliverables.