Meituan open-sources LongCat-2.5, a 1.6 trillion parameter agent model under MIT
Meituan, the Chinese food delivery and e-commerce company, placed the LongCat-2.5 checkpoint on Hugging Face on November 9 under the MIT License, according to <cite index="2-1">VentureBeat</cite>. The company had first introduced the model on <cite index="10-9">November 1, 2026</cite>, so the open release came about a week later. The system is a natively multimodal mixture-of-experts design that can inspect screenshots and video, work through a codebase, and write or revise software over several steps. Per the model card, <cite index="2-5">about 48 billion of its 1.6 trillion parameters are active per token</cite>, and the model is built for long-horizon agent work with a <cite index="12-2">1M context window</cite>.
Meituan's reported gains over LongCat-2.0 are large. Terminal Bench v2.1 rose to 88.4 from 70.8, <cite index="6-5">compared with 70.8 for its predecessor</cite>, and DeepSWE v1.1 climbed to 64.6 from 9.7. On the SWE Atlas codebase question-answering test, the score went from 33.1 to 58.3, as <cite index="1-7">VentureBeat reports</cite>. The comparisons against newer rivals are less favorable: on Terminal Bench v4.0, <cite index="6-7">LongCat-2.5 scored 38.5 while Claude Opus 5 scored 53.9</cite>. Limited-time hosted API rates are listed at <cite index="5-3">$0.30 per million uncached input tokens and $1.20 per million output tokens</cite>. Reporting also differs on size, since <cite index="6-11">the Hugging Face repository metadata displays 1.8T parameters, while Meituan's technical description states 1.6T</cite>.
Why it matters
The MIT License is the most consequential detail for product teams. It permits <cite index="2-4">commercial use, modification, and redistribution subject to its notice requirements</cite>, and it grants no rights to Meituan's trademarks or patents. That gives companies a path to self-host a frontier-scale agent model and fine-tune it on proprietary code or internal workflows instead of relying only on a vendor API.
The hardware bill is the constraint. <cite index="6-10">A uniform 16-bit representation of 1.6 trillion parameters would require about 3.2 TB for weights alone, before runtime buffers and context memory.</cite> Founders should run the model against their own repository tasks and compare serving cost with smaller open models before committing. For investors, the open release plus low hosted pricing puts pressure on the margins of closed API providers, while self-hosting still favors operators with large GPU budgets.
