Meituan LongCat-2.5: 1.6-trillion-parameter AI agent model

Meituan, the Chinese food-delivery and local-services giant, has released LongCat-2.5, the newest model from its LongCat AI team. The company presents it as a native multimodal model for long-horizon agent tasks: jobs where an AI has to keep working through many steps, such as building a whole software project. Details come from the official LongCat blog post and Meituan's announcement on 2 November.

The key specs

• Size: 1.6 trillion total parameters in a mixture-of-experts (MoE) design, with about 48 billion activated for each token. MoE models only use a fraction of their "experts" at a time, which keeps the compute per answer far below what the headline number suggests.
• Multimodal by design: images are turned into multi-level discrete tokens, so the same backbone processes pictures and text in one format. Meituan says this helps with "visual coding" (turning designs into working pages), creative design and 3D tasks.
• Built for agents: training added more long task sequences and large-scale reinforcement learning on real codebases. Meituan also says it converts failures by real agents into reproducible training tasks.
• Chinese hardware: both training and inference were optimised for domestic AI accelerators with less than 80 GB of memory each. The company describes a 4-bit (W4A4) inference setup that, in its stress tests, raised prefill throughput by 43.3% and decode throughput by 42.77% compared with an 8-bit baseline.

How it compares, by Meituan's numbers

The blog includes a large benchmark table against GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5 and other Chinese models. Results for rival models are partly taken from public reports and partly self-run, so treat them as indicative, not independent. A few examples:

• Terminal Bench v4.0 (agentic coding): LongCat-2.5 scores 38.5, behind Claude Opus 5 at 53.9 and GLM-5.3 at 41.8.
• BrowseComp (web research): 91.9, slightly above Claude Opus 5's 90.8.
• AutomationBench: 51.8, ahead of Claude Opus 5 (50.3) but behind DeepSeek-V4.1-Flash (54.8).

The picture is a strong agent model that is competitive on browsing and automation tasks, but not the leader in hard coding tests.

How to try it

LongCat-2.5 is available now through Meituan's LongCat chat service and its API platform. Meituan says the weights will follow on GitHub and Hugging Face, which would make it one of the largest open-weight models available. If that happens, keep expectations realistic: even with only 48B active parameters, a 1.6-trillion-parameter model needs data-centre hardware. Our guide on how much VRAM you need for local AI explains why, and our explainer on what open-weight AI models are covers what "open" does and doesn't allow.

Why it matters

Meituan is not an AI lab first, yet it now ships trillion-parameter models trained entirely on Chinese chips. That is another sign that US export limits on Nvidia hardware have not stopped Chinese companies from building frontier-scale systems. It also adds to a crowded field of Chinese agent models alongside DeepSeek, Kimi and GLM, all competing on price and openness.