美團開源一點六萬億參數代理模型權重,訓練全程不用英偉達芯片Meituan open-sources 1.6-trillion-parameter agent model trained without Nvidia chips
- 事件
- 美團以麻省理工許可開源 LongCat-2.5 權重,一點六萬億參數、每次激活四百八十億。稱訓練全用國產卡,指南列昇騰Meituan released the weights of LongCat-2.5 under the MIT licence, a mixture-of-experts model with 1.6 trillion total parameters activating about 48 billion per call, and a one-million-token context. It says training and fine-tuning ran entirely on Chinese accelerators, and its self-hosting guide lists only 194 Huawei Ascend cards.
- 背後
- 開源模型多在英偉達集群訓練,出口管制令算力溢價集中在少數廠商。權重與部署指南同時開放,等於把非英偉達推理路線交給企業。Open models have mostly been trained on Nvidia clusters, so export controls kept the frontier-compute premium with a few vendors. Releasing both weights and a deployment guide hands enterprises an inference alternative that does not run on Nvidia.
我們認為,這條消息真正的份量不在榜單分數,而在部署指南。當一家企業能在自家機房按圖施工,模型採購就由租用服務變成資本支出決策,議價基準亦會隨之改變。In our view the weight of this release is not its benchmark score but its deployment guide. Once a firm can follow a diagram inside its own server room, model procurement turns from a rental decision into a capital decision, and the benchmark for negotiation moves with it.
- 對日常工作的影響
- IT 主管須重新計算自建推理的總擁有成本,包括機房、電力與運維人手;採購部門則要多問一句,合約是按詞元計價還是按年授權。IT leads must recompute the total cost of self-hosted inference, including space, power and operations staff, while procurement adds one question: is the contract priced per token or per annual licence?
- 如何改善
- 本週挑一條流量穩定的內部問答流程,用開源權重在本機跑通一次,記下每千次調用的時間與電費,存檔作下一輪議價底稿。This week, run one steady internal Q&A workflow on the open weights locally, log the time and electricity cost per thousand calls, and file it as the baseline for your next negotiation.
