Meituan Open-Weights 1.6T LongCat-2.5 Agent Model Under MIT

Meituan, the Chinese food delivery and e-commerce platform, published the full open weights of LongCat-2.5 on Hugging Face on November 9, 2026, one week after introducing the model. The release carries an MIT license that permits commercial use and modification without restriction. The 1.8-terabyte checkpoint contains 1.6 trillion total parameters operating under mixture-of-experts routing, with roughly 48 billion parameters active per token during inference.

Release Date and Licensing Terms

Officially released on November 9, 2026, LongCat-2.5 became available under the MIT license, one of the most permissive open-source licenses in software. The move places the model in the small category of trillion-parameter systems where developers can download, modify, and deploy the weights for any purpose, including commercial applications. The full checkpoint weighs 1.8TB and requires significant storage and memory infrastructure for self-hosting.

Architecture and Context Window

LongCat-2.5 operates as a native multimodal mixture-of-experts model with a 1-million-token context window designed for long-horizon agent tasks. The 1.6T total parameter count reflects the full expert set, while routing activates approximately 48 billion parameters per token during generation. This architecture targets scenarios where models must maintain state across extended coding sessions, multi-step debugging workflows, or complex question-answering chains that exceed the capacity of standard context windows.

Benchmark Performance Against Frontier Models

The model posts 88.4 on Terminal Bench v2.1, 64.6 on DeepSWE v1.1, and 78.6 on Vision2Web according to Meituan's published benchmarks. These figures represent substantial gains over the prior LongCat release, which scored 70.8 on Terminal Bench v2.1 and 9.7 on DeepSWE v1.1. Against current closed frontier models, LongCat-2.5 lands within one point of GPT-5.6 Sol and Claude Opus 5 on Terminal Bench v2.1 but trails Claude Opus 5 on the v4.0 variant at 38.5 to 53.9 and falls behind DeepSeek-V4.1-Flash on DeepSWE at 74.2. Additional benchmarks include 58.3 on SWE Atlas codebase question-answering and 55.4 on 3DCodeBench.

Hosted API Pricing and Self-Hosting Options

Meituan is offering hosted inference during a limited-time launch window at $0.30 per million uncached input tokens and $1.20 per million output tokens. Cached input tokens drop to $0.006 per million. For teams that prefer self-hosting, Meituan published an SGLang-FluentLLM deployment guide targeting neural processing units, with four-bit precision optimization for expert weights aimed at Chinese accelerators. The hosted API provides a lower-friction entry point for developers who lack the infrastructure to run a 1.8TB model locally.

Procurement and Compliance Considerations

The release arrives amid growing scrutiny in Washington over Chinese AI models entering U.S. government and contractor workflows. VentureBeat flagged the need for suppliers to review contract terms before integrating open-source models from Chinese labs into government-facing work. LongCat-2.5 is the latest in a dense sequence of more than 300 China AI stories tracked over the past ninety days, many involving open-weights releases where compute costs are low but procurement and compliance overhead can become the primary friction point for enterprise buyers.

What This Means

LongCat-2.5 extends the trend of Chinese labs releasing trillion-parameter models under permissive licenses, closing the performance gap with Western frontier systems while offering significantly lower hosting costs. The MIT license removes legal barriers to modification and commercial use, but teams in regulated sectors will need to account for export control and procurement policy before deployment. For developers working on long-context agent applications outside those constraints, the combination of 1M-token context, competitive agent benchmarks, and $0.30 input pricing creates a viable alternative to closed API providers.