Meituan Open-Sources 1.6-Trillion-Parameter LongCat-2.5 Under MIT License

Meituan published the weights of LongCat-2.5, a mixture-of-experts model with 1.6 trillion total parameters, on Hugging Face under the MIT license on Nov. 9. The release came about a week after the Chinese food delivery and e-commerce group introduced the model. The license permits commercial use, modification and redistribution, but grants no rights to Meituan trademarks or patents.

The model activates roughly 48 billion parameters per token, according to Meituan's model card. The Hugging Face metadata lists 1.8 trillion total parameters, a discrepancy VentureBeat also flagged. LongCat-2.5 builds on LongCat-2.0 and handles text, images and video. It converts visual features into discrete tokens processed alongside text, and offers switchable thinking and non-thinking modes.

How does it score against Western rivals?

On Meituan's figures, LongCat-2.5 scores 88.4 on Terminal Bench v2.1. GPT-5.6 Sol scored 88.8 and Claude Opus 5 scored 89.1. DeepSeek-V4.1-Flash led at 90.6. Meituan marks some competitor figures as taken from public sources and calls its in-house numbers indicative only.

Other results from the model card:

• Terminal Bench v4.0: 38.5, against 53.9 for Claude Opus 5
• DeepSWE v1.1: 64.6, against 74.2 for DeepSeek-V4.1-Flash
• BrowseComp: 91.9, ahead of all three rivals
• HLE with tools: 63.1, against 64.7 for Claude Opus 5
• Vision2Web: 78.6, against 80.2 for Claude Opus 5

The largest gains are over its own predecessor. DeepSWE rose to 64.6 from 9.7 for LongCat-2.0. Terminal Bench v4.0 climbed to 38.5 from 3.0. On the harder Terminal Bench v4.0, LongCat-2.5 still trails Claude Opus 5 by more than 15 points. VentureBeat cautioned that public leaderboard results use different setups and are not a clean head-to-head.

Why four-bit quantization matters

Meituan quantized expert weights and activations to four bits, a W4A4 scheme, for domestic Chinese AI accelerators. The company reports 43.3% higher prefill throughput and 42.77% higher decode throughput than an eight-bit baseline. Deployment on NPUs runs through Meituan's SGLang-FluentLLM project.

The quantization work signals where the model is meant to run. Chinese developers face tightening access to U.S.-made accelerators, and W4A4 support targets the domestic NPU ecosystem directly.

Meituan's hosted API charges $0.30 per million input tokens and $1.20 per million output tokens at a limited-time rate, VentureBeat reported.

What is the regulatory picture?

As of Nov. 9, VentureBeat found no U.S. federal restriction naming Meituan or LongCat-2.5. That contrasts with the October order excluding DeepSeek from federal civilian procurement. The open weights remain downloadable and commercially usable in the United States for now.

The MIT terms, benchmark parity on several coding and browsing tasks, and native support for Chinese NPUs position LongCat-2.5 as a freely deployable rival to closed Western models, with the parameter-count discrepancy and uneven v4.0 results the clearest caveats for adopters.