Meituan releases LongCat-2.5, a 1.6-trillion-parameter model built for agentic tasks
In brief
- Meituan launched LongCat-2.5 on November 2, 2026, with access through an API.
- LongCat-2.5 is multimodal, with 1.6 trillion total parameters and about 48 billion active per token.
- Context window reaches one million tokens, with outputs of up to 128,000 tokens.
- Meituan claims the model is competitive with Claude Opus variants on several benchmarks.
- Domestically produced Chinese AI accelerators were used to train the model, Crypto Briefing reported.
What's under the hood
LongCat-2.5 uses approximately 48 billion active parameters per token (out of that 1.6 trillion total), per Meituan's specs as reported by Crypto Briefing. It supports a native context window of one million tokens and can produce outputs of up to 128,000 tokens. That's a lot of room for an agent working through a long task.
Images aren't bolted on as an afterthought. The model processes them as multi-level discrete tokens alongside text.
On training, Meituan used long-horizon reinforcement learning built around failure trajectories and synthetic data. For inference, the company applied W4A4 quantization, which uses 4-bit precision for both weights and activations, and Meituan reported substantial throughput improvements from the technique.
Built for agents, not chat
Meituan said the model's breadth makes it usable across terminals, browsers, graphical interfaces, spreadsheets and design tools. LongCat-2.5 builds on LongCat-2.0, which debuted in June 2026. The full release followed a preview on September 25, which showed capabilities like image understanding but didn't include performance benchmarks.
This time there are numbers. Meituan's own performance claims indicate LongCat-2.5 is competitive with Claude Opus variants on several benchmarking tasks.
Those are Meituan's claims, though.
The chip angle
Crypto Briefing reported that LongCat-2.5 was trained on domestically produced Chinese AI accelerators. The outlet tied that to Meituan's push for domestic technological independence, set against a tightening set of US export restrictions on advanced AI technologies (that framing is Crypto Briefing's, not a motive Meituan stated itself).
For a company best known for food delivery, it's a serious entry into the agentic AI race. The specs are big, the context window is long and the target use case is narrow on purpose: coding, automation and visual tasks.
Frequently asked questions
What is Meituan's LongCat-2.5?
LongCat-2.5 is a multimodal AI model from Chinese delivery company Meituan, launched November 2, 2026 and available through an API. It has 1.6 trillion total parameters, with about 48 billion active per token. Meituan built it for complex, long-horizon agentic tasks.
How large is LongCat-2.5's context window?
LongCat-2.5 supports a native context window of one million tokens. It can also produce outputs of up to 128,000 tokens, according to Meituan's specs as reported by Crypto Briefing.
Does LongCat-2.5 really rival Claude Opus?
Meituan's own performance claims indicate LongCat-2.5 is competitive with Claude Opus variants on several benchmarking tasks. These are company-reported results. The earlier September 25 preview didn't include performance benchmarks.
What hardware was LongCat-2.5 trained on?
Crypto Briefing reported that LongCat-2.5 was trained on domestically produced Chinese AI accelerators. The outlet linked this to Meituan's push for technological independence amid tightening US export restrictions on advanced AI technologies.
