Meituan releases LongCat-2.5 weights under the MIT licence

LongCat-2.5 is a native multimodal mixture-of-experts model built for long-horizon agent tasks, with 1.6 trillion total parameters and about 48 billion activated per token[1]. It builds on LongCat-2.0 and combines visual understanding with agentic coding for codebase work, full-project builds, design and 3D tasks[1].

The model turns native-resolution images into discrete tokens that sit in the same space as text, and it supports both a thinking mode and a faster non-thinking mode[1]. Meituan says expert weights and activations use 4-bit quantisation, which gave 43.3% higher prefill throughput and 42.77% higher decode throughput than an 8-bit baseline in its stress tests[1].

The biggest jumps are over its predecessor. Meituan reports 88.4 on Terminal Bench v2.1 against 70.8 for LongCat-2.0, 64.6 on DeepSWE v1.1 against 9.7, and 91.9 on BrowseComp against 79.9[1]. Most of those numbers are measured in-house, and Meituan labels them indicative only[1].

Against rivals the picture is mixed. VentureBeat points out that Meituan compared LongCat-2.5 with GPT-5.6 Sol and Claude Opus 5, whose successors were already out, and that newer models such as GPT-6.1 Sol and Claude Opus 5.5 post higher Terminal Bench v4.0 scores than LongCat's 38.5[2].

The MIT licence lets companies use, modify and redistribute the model commercially[2], which makes LongCat-2.5 one of the largest openly licensed multimodal agent models available. Running it is still a serious job: the full checkpoint has to be loaded even though only a slice is active per token, and Meituan's tested setup targets NPUs on Chinese accelerators[1]. For most teams, VentureBeat suggests a hosted trial first, where Meituan's limited-time rate is $0.30 per million input tokens and $1.20 per million output tokens[2].