觉
AI觉醒星球
Awakening is here
AI小生意项目库 · 阅读档案
2026-11-15 1 浏览 免费阅读

美团发布 1.6 万亿参数智能体模型,全程未使用英伟达 GPU

美团开源 LongCat-2.5:1.6 万亿参数 MoE 智能体模型,完全基于中国硬件训练,支持百万上下文与多模态输入,MIT 许可开放权重;部分编码与终端基准接近西方头部模型,但更难任务仍有差距。

AI小生意项目库 约 9 分钟读完 免费阅读 原文 2026-11-15 18:06:06

原贴

查看原文
作者:Jonathan Kemper 来源站点:the-decoder.com 原贴时间:

原文

Meituan releases the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents. It was trained entirely on Chinese hardware and performs at the level US models like Claude Opus 5 reached a few months ago on coding benchmarks. The LongCat team at Chinese food delivery company Meituan has released the weights for LongCat-2.5, one week after the model's official unveiling. LongCat-2.5 is built for agents that work through long, multi-step tasks autonomously, particularly coding. According to Meituan, it comes close to top Western models on several benchmarks. LongCat-2.5 is a mixture-of-experts model that activates only a subset of its specialized subnetworks for each token it processes. Out of 1.6 trillion total parameters, roughly 48 billion are active per token. Through the API, the model handles up to one million tokens at once and generates responses up to 128,000 tokens long. It understands text, images, and video, responds in text, and lets users toggle reasoning on and off. LongCat-2.5 builds on its predecessor LongCat-2.0 and adds image understanding. Many multimodal models bolt on a separate image encoder after the fact. LongCat-2.5 instead converts images into tokens that the model processes alongside text, as is standard practice for today's LLMs. The approach draws on the research model LongCat-Next and aims to improve visual coding, design, and 3D tasks. For reinforcement learning, the team used real software tasks from GitHub pull requests along with recreated agent failures. To make sure the model actually solved the tasks during training, the developers blocked internet access in the training environments, removed Git history and cached files, and had a language model check its solutions for tricks. The team also distilled the capabilities of several specialized teacher models for agent work, reasoning, and dialog into LongCat-2.5. Meituan quantizes the experts' weights and activations to 4-bit precision, which Chinese chips can process natively. In Meituan's load tests, throughput rises 43.3 percent for prompt processing and 42.77 percent for response generation compared with an 8-bit version. Meituan says accuracy stays on par with the uncompressed model. This version is available in a separate repository . The self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series. Two handle image processing, 64 handle prompt ingestion, and 128 handle response generation. There's no guidance for Nvidia GPUs. The LongCat team reports 88.4 points on Terminal Bench v2.1, a benchmark that measures how well agents solve command-line tasks. Meituan lists Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8, and Deepseek-V4.1-Flash at the top of the table with 90.6. DeepSWE tests how well models fix bugs in real code repositories. Here LongCat-2.5 jumps from 9.7 to 64.6 points, placing it behind every other model in the table except Claude Opus 4.8 , with Deepseek-V4.1-Flash leading at 74.2 and GPT-5.6 Sol at 72.7. The more recent Claude Opus 5.5 and GPT-6.1 Sol are not included in the table. Terminal Bench v4.0, the harder variant, puts LongCat-2.5 at 38.5 points, slightly above GPT-5.6 Sol's 37.3 but well below Opus 5's 53.9. The model posts 91.9 on the web search benchmark BrowseComp, the highest score in the table, though Meituan ran this test in its own environment. The weights are available under the MIT license, which allows commercial use, modification, and redistribution. Users can also try LongCat-2.5 in a chat interface at https://longcat.ai/ . Its API supports both the OpenAI and Anthropic formats and plugs into agent harnesses like Claude Code, Codex, and Hermes. At launch, Meituan is charging a limited-time promotional price equivalent to roughly €0.27 per million input tokens and €1.07 per million output tokens. Meituan says it trained the predecessor LongCat-2.0 in late June entirely on a cluster of more than 50,000 Chinese AI chips . Before that, the mode

中文翻译

美团发布了 LongCat-2.5 的权重,这是一个为 AI 智能体构建的 1.6 万亿参数模型。它完全在中国硬件上训练,在编码基准测试中的表现达到美国模型如 Claude Opus 5 几个月前达到的水平。

核心信息

美团开源 LongCat-2.5:1.6 万亿参数 MoE 智能体模型,完全基于中国硬件训练,支持百万上下文与多模态输入,MIT 许可开放权重;部分编码与终端基准接近西方头部模型,但更难任务仍有差距。

  • 美团开源 LongCat-2.5:1.6 万亿参数 MoE 智能体模型,完全基于中国硬件训练,支持百万上下文与多模态输入,MIT 许可开放权重;部分编码与终端基准接近西方头部模型,但更难任务仍有差距。
  • 原贴提到:Meituan releases the weights for LongCat-2.5, a 1.6-trillion-parameter m
  • 来源:the-decoder.com

详细解读

这是什么信号:美团把 LongCat-2.5 的权重开放出来,这是一个为智能体任务构建的 1.6 万亿参数 MoE 模型,每个 token 约激活 480 亿参数,支持最高 100 万 token 上下文和 12.8 万 token 输出,能处理文本、图像、视频并以文本回复,推理开关可调。更关键的是,原文称它完全在中国硬件上训练,自托管指南只描述华为昇腾 Atlas A2 集群,没有英伟达 GPU 指引;MIT 许可允许商用、修改和再分发。这不是单点模型发布,而是中国芯片—训练—推理—开源许可的一条链路信号。

为什么重要:智能体模型的竞争点正在从聊天能力转向长程任务、编码、工具调用和可自托管。LongCat-2.5 在 Terminal Bench v2.1 拿到 88.4,接近原文列出的 Claude Opus 5 的 89.1 和 GPT-5.6 Sol 的 88.8,低于 Deepseek-V4.1-Flash 的 90.6;但在更难的 Terminal Bench v4.0 只有 38.5,低于 Opus 5 的 53.9;DeepSWE 从 9.7 提升到 64.6,仍落后 Deepseek 和 GPT 的表中成绩。结论是:它证明国产硬件可训练 frontier 级智能体模型,并在部分任务接近头部,但还不是全面领先。原文还强调 4-bit 量化让中国芯片原生处理,吞吐相较 8-bit 版本提升 43.3% 和 42.77%,且美团称精度不降——这比单纯堆参数更有产业意义。

对谁有价值:第一,做编码智能体、终端自动化和长上下文工具链的团队,可以用 MIT 权重做私有化部署或二次训练,并通过 OpenAI、Anthropic 兼容 API 接入 Claude Code、Codex、Hermes 等 harness。第二,国产算力与推理服务商,可把“昇腾集群 + 4-bit 量化 + 长上下文”作为方案样板。第三,企业架构师和投资/研究者,可用它观察中国 AI 全栈能力与开源模型竞争格局。第四,成本敏感型业务可把它作为第二供应商,降低对单一海外模型的依赖。

可以怎么行动:先不要只看榜单。用自己真实的代码库、Issue、CI 日志和终端任务做小规模 A/B,重点测长程多步任务、工具调用稳定性、图像理解、百万上下文下的召回与延迟。技术上验证 OpenAI/Anthropic 格式兼容、agent harness 接入成本、4-bit 版本精度损失和 194 卡昇腾集群的部署门槛。商业上把它放进多模型路由:高价值任务用最强模型,批量编码与私有数据任务用开源权重兜底。采购上要求供应商提供独立复现的 Terminal Bench、DeepSWE 结果,而不是只接受自报数据。

风险或限制:原文的基准多为美团自报,BrowseComp 的 91.9 还是在其自有环境跑出,且表中未包含更新的 Claude Opus 5.5 与 GPT-6.1 Sol;更难的 Terminal Bench v4.0 和 DeepSWE 显示它与头部仍有明显差距。1.6 万亿总参数即使只激活 480 亿,训练与推理成本、显存和集群调度仍不低;自托管指南绑定昇腾 A2,缺少英伟达路径会限制海外或现有 GPU 团队采用。4-bit 精度“不降”的说法需要独立验证;MIT 许可虽宽松,但数据合规、安全对齐、出口管制和国产芯片供应仍是落地变量。对智能体而言,训练时用 GitHub PR 和复现失败任务、断网、去 Git 历史、用语言模型查作弊,说明团队重视可靠性,但这不等于生产环境零风险。

信息差价值

这条内容的真正价值,不只是“有人发布了一个新功能”,而是它揭示了 the-decoder.com 背后的产品方向、工作流变化或竞争信号。对 OPC 来说,这种信息可以转化成持续追踪的栏目选题。

如果把《美团发布 1.6 万亿参数智能体模型,全程未使用英伟达 GPU》放到你的内容系统里,它最大的价值在于帮助读者更快看懂“为什么值得关注”,而不是只看到一条碎片化动态。

参考来源

AI SUMMARY

这篇文章回答了什么

美团发布 1.6 万亿参数智能体模型,全程未使用英伟达 GPU主要讲什么?

美团开源 LongCat-2.5:1.6 万亿参数 MoE 智能体模型,完全基于中国硬件训练,支持百万上下文与多模态输入,MIT 许可开放权重;部分编码与终端基准接近西方头部模型,但更难任务仍有差距。

这篇文章最值得关注的要点是什么?

美团开源 LongCat-2.5:1.6 万亿参数 MoE 智能体模型,完全基于中国硬件训练,支持百万上下文与多模态输入,MIT 许可开放权重;部分编码与终端基准接近西方头部模型,但更难任务仍有差距。;原贴提到:Meituan releases the weights for LongCat-2.5, a 1.6-trillion-parameter m;来源:the-decoder.com

这篇文章和哪些AI专题相关?

它适合放在AI副业、Agent工作流、AI工具专题里阅读。 关联原因:这篇内容命中「项目、小生意、变现」等主题信号。;这篇内容命中「智能体」等主题信号。;这篇内容命中「模型」等主题信号。

阅读这篇文章建议先理解哪些关键词?

建议先理解AI工具、工具、自动化、模型、Cursor这些关键词,再结合正文判断工具、机会或风险是否值得进入自己的工作流。

上一篇 langchain-xai 1.4.0 发布:ChatXAI 改为通过 LangSmith 网关路由 下一篇 Anthropic CEO Amodei 承认“明年会发生什么存在巨大不确定性”,就在创纪录 IPO 前几天
北竹游乐场 免费玩小游戏 免费玩