Chinese technology platform and consumer services giant Meituan officially announced the comprehensive open-source release of its flagship frontier multimodal mixture-of-experts (MoE) foundation model, LongCat-2.5, on November 10, 2026. Distributed under the commercial-friendly MIT License, the model features an aggregate capacity of 1.6 trillion parameters while dynamically routing approximately 48 billion active parameters per token during inference. LongCat-2.5 natively supports an extensive 1-million-token context window engineered specifically for long-horizon autonomous software agent workflows.
LongCat-2.5 architecture and benchmark scores of 88.4 and 64.6
The model architecture incorporates native multi-modal cross-attention layers trained directly across interleaved high-resolution imagery, complex code repositories, and sequential natural language text.
By embedding dynamic positional rotary embeddings alongside sparse attention optimizations during pre-training, LongCat-2.5 maintains stable associative recall across million-token sequences without memory degradation. Announcing the open-weights distribution during a developer presentation, a senior Meituan Artificial Intelligence research director stated: "Building capable autonomous agent systems requires models with persistent long-term memory and deterministic multi-step tool execution. LongCat-2.5 couples trillion-parameter reasoning capacity with high inference throughput under the MIT License, empowering developers worldwide to deploy commercial-grade agent architectures without restrictive licensing barriers."
Standardized benchmark evaluations demonstrated competitive performance across rigorous reasoning and code generation tasks, with LongCat-2.5 securing an 88.4 score on Terminal Bench v2.1 and 64.6 on DeepSWE v1.1.
Meituan releases 1.8-terabyte checkpoints for enterprise self-hosting
The benchmark achievements approach the operational accuracy of leading closed-source commercial foundation models across extended agent execution loops. To facilitate practical enterprise self-hosting, Meituan released pre-quantized 1.8-terabyte model checkpoints optimized for distributed deployment using SGLang and vLLM inference engines. Commercial developers can execute high-throughput on-premises model hosting across two standard multi-GPU server nodes, significantly mitigating cloud API dependency expenses for sensitive enterprise applications.
Artificial intelligence open-source researchers commended Meituan's decision to distribute open weights under permissive licensing, emphasizing that open trillion-parameter MoE architectures democratize access to advanced autonomous reasoning infrastructure across global scientific and commercial developer communities.
Artificial intelligence policy scholars and open-source software engineers emphasized that distributing trillion-parameter mixture-of-experts models under permissive open-source licenses fundamentally accelerates scientific innovation. Because developing frontier foundation models requires immense capital investments in high-performance computing clusters, releasing open weights empowers researchers, academic institutions, and enterprise developers to construct specialized domain-specific agent architectures without recurring API licensing overhead. Computer science faculties commended the native multi-modal architecture, noting that million-token context windows support transformative automation workflows.
Furthermore, machine learning practitioners highlighted that comprehensive multilingual pretraining benchmarks demonstrate robust reasoning capabilities across diverse coding and scientific tasks. Open-source developer communities look forward to deploying customized fine-tuned models for specialized domain applications.
In addition, machine learning researchers and software architecture specialists emphasized that releasing state-of-the-art trillion-parameter mixture-of-experts models under open-source licenses democratizes cutting-edge foundational technology for global developer communities. Academic institutions and independent engineering teams can conduct comprehensive safety audits, fine-tune domain-specific reasoning agents, and optimize inference pipelines without incurring restrictive proprietary API fees. Furthermore, open-source technology advocates commended the native multimodal and long-context capabilities for establishing new benchmarks across code synthesis, mathematical proofs, and document analysis.
Furthermore, open-source AI community organizers praised the transparent release of detailed pretraining methodologies, model architecture schematics, and reproducible benchmark evaluation suites. Software engineering teams worldwide look forward to building fine-tuned domain-specific reasoning agents and specialized developer tools powered by the LongCat-2.5 open multimodal foundation model.
Open-source AI architectures empower global researchers and developers
In conclusion, artificial intelligence research foundations commended the engineering achievement, noting that open foundational models drive collaborative global progress. Developers worldwide look forward to creating impactful machine learning solutions that benefit society at large. Open source AI architectures will continue to empower researchers across international universities and software development laboratories worldwide.
