{"slug": "meituan-s-longcat-2-5-jumps-into-china-s-open-ai-coding-race", "title": "Meituan's LongCat-2.5 Jumps Into China's Open AI Coding Race", "summary": "Meituan released LongCat-2.5 on November 1, a 1.6-trillion-parameter mixture-of-experts coding model that scored 88.4 on Terminal-Bench v2.1 (up from 70.8 for LongCat-2.0) and 64.6 on DeepSWE v1.1 (up from 9.7). The model activates about 48 billion parameters per task, handles a million-token context window with up to 128,000 output tokens, and is natively multimodal, scoring 78.6 on Vision2Web and 55.4 on 3DCodeBench's text-to-3D task. LongCat-2.5 is API-only at a limited-time rate of $0.30 per million uncached input tokens, with GitHub and Hugging Face releases promised but weights not yet published, unlike LongCat-2.0, which shipped under an MIT license.", "body_md": "*A food delivery giant just shipped a 1.6 trillion parameter coding model that beats its own predecessor by double digits on agentic benchmarks, and it isn't done yet.*\n\nMeituan released LongCat-2.5 on November 1, its newest large language model built for long, multi-step coding and design work. The jump from its own prior model is enormous. On Terminal-Bench v2.1, a benchmark that tests an AI agent's ability to complete real command-line tasks, LongCat-2.5 scored 88.4, up from 70.8 for LongCat-2.0. On DeepSWE v1.1, a tougher agentic coding benchmark, it posted 64.6 against LongCat-2.0's 9.7. That's a nearly sevenfold gain in under three months.\n\nMeituan is China's largest food delivery company, the one that sends scooters weaving through Beijing traffic to drop off hot pot and bubble tea. It is also, somewhat quietly, building one of the more serious open-weight model labs in the country under its LongCat brand, run separately from the consumer app. LongCat-2.5 is a mixture-of-experts model with 1.6 trillion total parameters, of which only about 48 billion activate for any given task. It's the same scale as its predecessor but retrained for far longer task chains. It handles a million-token context window. It can also output up to 128,000 tokens in a single response, which matters for a model meant to work through an entire codebase rather than one function at a time.\n\nThe model is natively multimodal, meaning it processes text and images through the same underlying token system instead of bolting a separate vision module on top. That lets it take on what Meituan calls visual coding. It can build a working interface from a screenshot or a sketch, handle 3D asset generation, or reason about a webpage's layout directly from a rendered image rather than its HTML. On Vision2Web, a benchmark measuring how well a model can translate a visual mockup into functioning code, LongCat-2.5 scored 78.6. On 3DCodeBench's text-to-3D task, it scored 55.4.\n\nLongCat-2.5 is, for now, API-only. Meituan has said GitHub and Hugging Face releases are coming, but hasn't published weights yet, which puts the newest model in an odd spot next to its own predecessor. LongCat-2.0 shipped under an MIT license over the summer, and developers have already been running it through Codex CLI and other agent harnesses as a genuinely open alternative to closed frontier coders. Until Meituan actually posts the LongCat-2.5 weights, that openness claim belongs to the older model, not the one making headlines this week.\n\n[Revolut Says Self-Hosted Open AI Models Cut Its Coding Costs 90%](https://startupfortune.com/revolut-says-self-hosted-open-ai-models-cut-its-coding-costs-90/)\n\nRevolut's infrastructure lead, David Abitbol, detailed the switch at GTC Berlin on October 22, saying the fintech now pays roughly 30 cents per million tokens, down 90%, while image tasks still route to Gemini. - [self-hosted open source models reduce coding costs](https://startupfortune.com/revolut-says-self-hosted-open-ai-models-cut-its-coding-costs-90/) - [open AI model deployment saves engineering expenses](https://startupfortune.com/revolut-says-self-hosted-open-ai-models-cut-its-coding-costs-90/)\n\nThat gap matters because the competition isn't waiting. Alibaba's Qwen team has been shipping aggressively all year: Qwen3-Coder launched back in June with Apache 2.0 licensing, and Qwen3.6 has since pushed its SWE-Bench Verified score to 73.4, ahead of Google's Gemma 4. DeepSeek, the other fixture of China's open-weight scene, keeps its own coding variants in regular rotation. None of these labs is waiting for permission from Washington or Silicon Valley to compete on agentic coding. Meituan's entry adds a fourth serious name, a food delivery company turned frontier-adjacent lab, to a list that a year ago was mostly DeepSeek and Alibaba.\n\nMeituan is also pricing LongCat-2.5 aggressively through its API. The limited-time rate is $0.30 per million uncached input tokens, just $0.006 per million for cached input, and $1.20 per million output tokens. Those are the kind of numbers that make a developer building an agent pipeline pay attention regardless of which lab's logo is on the model card.\n\nFor developers choosing between Qwen, DeepSeek and LongCat right now, the honest answer is that LongCat-2.5's biggest selling point, the open weights, doesn't exist yet. Its benchmark gains over LongCat-2.0 are real and large. But Meituan published no head-to-head numbers against Qwen4 Coder, Qwen3.6 or DeepSeek's latest. So the comparison a developer actually wants, apples to apples against the models they'd otherwise pick, isn't available from the company itself. What is available is a cheap, capable API and a promise that the weights are coming. In a market this crowded, that promise is worth less than the file itself.\n\n**Also read:** [Revolut Says Self-Hosted Open AI Models Cut Its Coding Costs 90%](https://startupfortune.com/revolut-says-self-hosted-open-ai-models-cut-its-coding-costs-90/) • [Morgan Stanley CEO Says AI Buildout Could Hit $15 Trillion and We're Barely Started](https://startupfortune.com/morgan-stanley-ceo-says-ai-buildout-could-hit-15-trillion-and-were-barely-started/) • [Apple guides margins lower again as memory chip costs keep climbing](https://startupfortune.com/apple-guides-margins-lower-again-as-memory-chip-costs-keep-climbing/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/meituan-s-longcat-2-5-jumps-into-china-s-open-ai-coding-race", "canonical_source": "https://startupfortune.com/meituans-longcat-25-jumps-into-chinas-open-ai-coding-race/", "published_at": "2026-11-03 06:58:10+00:00", "updated_at": "2026-11-03 07:22:38.940136+00:00", "lang": "en", "topics": ["large-language-models", "generative-ai", "ai-research", "ai-tools", "artificial-intelligence"], "entities": ["Meituan", "LongCat-2.5", "LongCat-2.0", "Alibaba", "Qwen3-Coder", "DeepSeek", "Revolut", "David Abitbol"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/meituan-s-longcat-2-5-jumps-into-china-s-open-ai-coding-race", "markdown": "https://wpnews.pro/news/meituan-s-longcat-2-5-jumps-into-china-s-open-ai-coding-race.md", "text": "https://wpnews.pro/news/meituan-s-longcat-2-5-jumps-into-china-s-open-ai-coding-race.txt", "jsonld": "https://wpnews.pro/news/meituan-s-longcat-2-5-jumps-into-china-s-open-ai-coding-race.jsonld"}}