--- id: "2026-09-25-meituan-longcat-2-5-preview" url: "https://postcutoff.com/e/2026-09-25-meituan-longcat-2-5-preview/" as_of: "2026-10-14T00:22:00+02:00" date: "2026-09-25" date_precision: day category: model-release importance: 3 confidence: medium status: [Partly confirmed] sources: 4 editor: human editor human_review: null version: "2026-10-13" --- As of: 2026-10-14 00:22 CEST. Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/ Canonical page: https://postcutoff.com/e/2026-09-25-meituan-longcat-2-5-preview/ # Meituan puts LongCat-2.5-Preview on its API Full title: Meituan puts LongCat-2.5-Preview on its API: a 1.6T-parameter MoE with image understanding, 1M context and $0.30 / $1.20 pricing; it becomes a top-10 model on OpenCode On Sept 25, 2026 Meituan's LongCat team released LongCat-2.5-Preview on its API platform and web app: a mixture-of-experts model with about 1.6 trillion parameters (about 48B active), a 1M-token context window and new image understanding, built for long agentic coding tasks and compatible with Claude Code, OpenCode, Kilo Code, OpenClaw and Hermes. Meituan published no benchmark scores and no weights. Aggregator listings price it at $0.30 / $1.20 per 1M tokens; OpenCode's public data ranks it #9 by tokens in the week to Oct 13. Its predecessor LongCat-2.0 (June 30, 2026) was billed as the largest model trained on Chinese-made AI chips. ## Key facts - Released Sept 25, 2026 on the LongCat API open platform with a web experience (IT Home, via 10jqka, Sept 26) - Size: about 1.6 trillion parameters, mixture-of-experts, about 48 billion active per step; natively supports a 1M-token context window (IT Home). Max output about 131K tokens (OpenCode data; a search summary of other coverage says 128K) - New over LongCat-2.0, per the platform changelog as reported: image understanding (cross-modal Q&A, visual reasoning), stronger coding, and compatibility with Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code - Price (OpenCode model data, Oct 13): $0.30 input / $1.20 output / $0.006 cached input per 1M tokens; reasoning and tool calling supported; inputs text and image; open weights: no - No benchmarks: Chinese coverage noted that Meituan published no scores with this release (search summary of aitoollab.cn; not read directly) - Usage (OpenCode public data, updated Oct 13, 2026): #9 model by tokens in the last week (2.1T tokens); 6.2T tokens since Sept 25, 172K unique users, 8.8M sessions; daily volume peaked near 545B tokens on Sept 30. Recorded spend is $0, i.e. it was free to use through OpenCode in this period. Top countries by tokens: United States 10.8%, China 9.4%, India 7.6% - Background, LongCat-2.0 (June 30, 2026): a 1.6T-parameter MoE that Meituan said was trained on domestic AI ASIC infrastructure rather than Nvidia hardware (SCMP headline: 'China claims biggest AI model trained on local chips'); announced as open source with weights 'coming soon' (repository meituan-longcat/LongCat-2.0). Reports identified OpenRouter's anonymous 'Owl Alpha' as LongCat-2.0-Preview (runtimewire summary; primary sources not read) - Not verified: the official API endpoint (third-party pages disagree between api.longcat.ai and api.longcat.chat), the knowledge cutoff, and whether 2.5 was also trained on domestic chips ## What happened Meituan, the Chinese delivery and local-services company, runs a model team called LongCat. Its LongCat-2.0, announced on June 30, 2026, was presented as a 1.6-trillion-parameter model trained on Chinese-made AI chips. On Sept 25 the team put a successor preview on its API: same size class, a 1M-token context, image input added, and tuning for long agentic coding sessions inside third-party harnesses such as Claude Code and OpenCode. Meituan gave no benchmark numbers. The clearest outside signal is usage: OpenCode's public statistics show the model taking about 2 trillion tokens a week within three weeks of release, ninth among all models on that platform, while it was free there. This entry is a backfill written on Oct 13, 2026 from IT Home's report (as republished by 10jqka), OpenCode's data page and a secondary summary of the LongCat-2.0 launch. Meituan's own pages were not read, so confidence is medium. ## Why it matters It is another sign that large Chinese companies outside the usual AI labs are shipping very large, very cheap models aimed at coding agents, and that developers outside China use them when the price is near zero. The domestic-chip training claim for the LongCat line, if it holds for 2.5, bears on how much US export controls constrain Chinese training runs. ## Your AI and this story - GPT-6 Astra (training cutoff April 2026): 148 days after its cutoff - Claude Opus 5.5 (training cutoff June 2026): 87 days after its cutoff - Gemini 4 Argon (training cutoff March 2026): 178 days after its cutoff - Grok 4.7 (training cutoff May 2026): 117 days after its cutoff ## Sources 1. [OpenCode Data: meituan/longcat-2.5-preview (usage, price, limits)](https://opencode.ai/data/meituan/longcat-2-5-preview.md) (opencode.ai, docs) 2. [LongCat: Introducing LongCat-2.0 (June 30, 2026; not read directly)](https://longcat.chat/blog/longcat-2.0/) (longcat.chat, official) 3. [IT Home via 10jqka: Meituan launches LongCat-2.5-Preview: 1.6T parameters, long-horizon tasks and multimodal (Sept 26, 2026)](https://stock.10jqka.com.cn/20260926/c680285995.shtml) (stock.10jqka.com.cn, press) 4. [Runtimewire: Meituan's LongCat team releases 1.6T-parameter AI model trained on domestic ASICs](https://www.runtimewire.com/article/meituan-longcat-2-0-domestic-asic-ai-model) (runtimewire.com, press) ## Changes - 2026-10-13 (filed): Created as a review ## Related - 2026-10-06: [Mistral releases Mistral Large 4 ("le Chonk")](https://postcutoff.com/e/2026-10-06-mistral-large-4/index.md) - 2026-10-05: [Reflection AI unveils Beam, a 501B-parameter Apache-2.0 open-weight MoE](https://postcutoff.com/e/2026-10-05-reflection-beam-501b-open-weight/index.md) - 2026-09-17: [Z.ai says GLM-5.3 largely built the inference stack that serves GLM-5.3-Flash, calling it an early step toward recursive self-improvement](https://postcutoff.com/e/2026-09-17-zhipu-glm-infra-agent-rsi/index.md) - Models: [LongCat-2.5-Preview](https://postcutoff.com/m/longcat-2-5-preview/) (`longcat-2-5-preview`)