Meituan launches LongCat-2.5: 1.6-trillion-parameter agent model that sees images

What it's about

Meituan, best known in China as a food delivery service and local services platform, unveiled LongCat-2.5 on November 2, 2026. It's a natively multimodal Mixture-of-Experts (MoE) model, in which only part of the model's building blocks, the "experts", are activated for each token. In total it has 1.6 trillion parameters, about 48 billion of which are active per token. It's available through the LongCat website and an API. A preview came out on September 25, 2026, but without benchmarks. The model builds on LongCat-2.0 from June 2026.

Built for long agent tasks

Meituan is targeting "long-horizon" agent work: the AI is meant to plan, use tools, carry out multiple steps, check results and keep going until a larger goal is reached. Instead of short code snippets, it's supposed to work through an entire software project, find bugs and make changes across multiple files. The native context window is one million tokens, with output of up to 128,000 tokens. That's meant to keep terminals, browsers, graphical interfaces, spreadsheets and design tools in working memory over many steps. According to the changelog, the model works with agent environments such as Claude Code, Codex and Hermes. According to Cryptobriefing, general chatting isn't the main focus.

Image understanding and training

The model represents images as multi-level discrete tokens (an image is broken into units that are processed like words) and handles them in the same system as text. Meituan cites uses such as visual coding (web pages and interfaces built from visual specs), creative design, 3D demos and visual reasoning. Demos included interactive fluid simulations, 3D exhibitions and miniature cities.

For training, Meituan relies on reinforcement learning (learning through reward) over long interaction chains, with data synthesis, environment building and training across different agent frameworks. One notable detail: when a real agent fails, the failure is turned into a reproducible training task with verifiable tests. Mid-training used sequences of up to one million tokens.

Efficiency on domestic hardware

According to Meituan, training and serving ran on AI accelerators produced in China. In a long-sequence test, the optimized training system reached 86% more throughput per device than a DDP baseline (Distributed Data Parallel, a standard method for distributed training). Inference uses W4A4 quantization: weights and activations are stored in just 4 bits. Compared with W8A8 INT8, Meituan reports the following in online tests:

These figures come from Meituan and haven't been independently verified.

Benchmarks and availability

Cryptobriefing reports that Meituan sees the model on par with Claude Opus variants on several benchmarks. The sources give no concrete numbers, and these are vendor claims. There are no open weights yet; according to the announcement, GitHub and Hugging Face releases are to follow. Pricing hasn't been disclosed. Meituan also uses its models internally, for example in the business assistant Miaoshou, which helps merchants with analytics, store management and customer acquisition.

Our take

LongCat-2.5 shows that Chinese companies are increasingly training large agent models on domestic hardware, against the backdrop of US export restrictions. The approach of using agent failures as training material is interesting. Much remains open: all performance claims come from the vendor, independent tests of coding and multimodality are missing, and so is pricing. You can't run the model locally right now. Even after a weights release, at 1.6 trillion parameters it would only run on server hardware.