China's food delivery giant just caught up with Claude on homegrown chips

Meituan, China's food delivery giant, released a new large model today, LongCat-2.5, and on benchmarks it is breathing down Claude's neck. The model has 1.6 trillion parameters, with about 48 billion active per token. The part the company is proudest of: the whole training run happened on Chinese accelerators, no foreign hardware involved.

LongCat-2.5 is built for long agent jobs: digging through huge codebases, building whole projects, turning an image into a website or a 3D page. It is natively multimodal, so images and text go through the model as one stream of tokens. Context is up to a million tokens, with up to 128 thousand tokens of output.

By Meituan's own numbers, LongCat-2.5 lands between Claude Opus 4.8 and Claude Opus 5 on most tests and beats Opus 5 on some. On AutomationBench it scores 51.8 against 50.3, on BrowseComp 91.9 against 90.8, on 3DCodeBench (text to 3D) 55.4 against 54.3. On Terminal Bench 4.0, though, the gap is wide: 38.5 against 53.9. One caveat matters: the comparison is with Opus 5 and GPT-5.6 Sol, not the very latest models from Anthropic and OpenAI.

There is a speed trick too. The model runs in W4A4, with both weights and activations squeezed to 4 bits. Compared with the 8-bit version, the company says throughput went up about 40%, the first token arrives 16% sooner and each next token 20% faster, with no quality loss on the tasks it checked.

You can try it right now at longcat.ai, and the API is open for developers. It works inside Claude Code, Codex and other agent harnesses.

Meituan CEO Wang Xing put it bluntly back in the summer: "We will not become a token factory." The models are mainly there to serve the company's own business, and LongCat already powers an AI assistant for restaurants and shops selling through Meituan.

Bottom line: in China even food delivery companies now build models at last year's Claude flagship level, and on homegrown chips.