Photo: U.Lucas Dubé-Cantin / Pexels The Chinese delivery giant's latest AI model targets coding and automation, claims parity with Claude Opus on several benchmarks, and was trained on domestic chips
Meituan, the Chinese company best known for getting dumplings to your door in under an hour, has released LongCat-2.5. It is a multimodal AI model with 1.6 trillion total parameters, now available through an API.
What Meituan actually shipped #
LongCat-2.5 officially launched on November 2, 2026. Meituan built it for complex, long-horizon agentic tasks, meaning multi-step jobs where the AI acts on its own rather than answering one question at a time.
The headline figure is 1.6 trillion total parameters. Parameters are the adjustable internal values a model learns during training. More of them generally means more capacity to store patterns and knowledge.
The more practical number is the active count. LongCat-2.5 uses approximately 48 billion active parameters per token. A token is a small chunk of text or data the model processes.
The model supports a native context window of one million tokens. It can also produce outputs of up to 128,000 tokens.
Meituan says that breadth makes the model usable across terminals, browsers, graphical interfaces, spreadsheets, and design tools.
The upgrades under the hood #
LongCat-2.5 builds on LongCat-2.0, which debuted in June 2026. The jump between the two comes down to two main areas.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The first is vision. The new model processes images as multi-level discrete tokens alongside text. Instead of treating a picture as a separate attachment, it breaks images into units the model can reason about in the same stream as words.
The second upgrade is training. Meituan used long-horizon reinforcement learning built around failure trajectories and synthetic data.
Meituan also optimized inference with W4A4 quantization. Quantization compresses the numbers a model uses internally, in this case down to 4-bit precision for both weights and activations. Meituan reports substantial throughput improvements from the technique.
Benchmarks, and a preview that didn’t have any #
The full release followed a preview on September 25, 2026. That earlier version showed off capabilities like image understanding, but it arrived without performance benchmarks.
The November launch fixed that. Meituan’s performance claims indicate LongCat-2.5 is competitive with Claude Opus variants on several benchmarking tasks.
Meituan is positioning LongCat-2.5 squarely around agentic coding and visual tasks. General conversation is not the main pitch.
The chip story #
One detail stands apart from the specs. LongCat-2.5 was trained on domestically produced Chinese AI accelerators.
That fits Meituan’s push for domestic technological independence. The backdrop is a tightening set of US export restrictions aimed at advanced AI technologies.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.