• 亿邦公众号
    • APP下载

美团开源1.6万亿参数智能体模型 全程未用英伟达GPU

作者:亿邦AI 文章来源:亿邦动力

2026-11-16 09:40

作者:亿邦AI 文章来源:亿邦动力

2026-11-16 09:40

邦小白快读

EN
全文速览

总1:核心信息:美团开源LongCat-2.5,总参数1.6万亿,专为长链路多步骤智能体设计,尤其面向编程,训练全程用国产芯片,未用英伟达GPU。

1. 混合专家架构,每token仅激活约480亿参数;API一次最多处理100万token,单次响应最长128000 token,支持文本图像视频输入、文本输出,可开关推理模式。

2. 图像直接转token与文本同管线,目标是提升视觉编程、设计及三维任务表现。

总2:实操。

1. 可在longcat.ai试用,API兼容OpenAI与Anthropic,可接Claude Code、Codex、Hermes;权重MIT许可,可商用修改再分发,限时价每百万输入token约0.27欧元、输出1.07欧元。

2. 自托管量化版有独立代码仓库,指南用194块华为昇腾Atlas A2,2块图像、64块提示摄取、128块响应生成;4比特量化使提示处理吞吐量提升43.3%、响应生成提升42.77%,官方称精度与未压缩模型持平。

3. 基准上Terminal Bench v2.1得88.4,DeepSWE从9.7升至64.6,但v4.0仅38.5,低于Opus 5的53.9,使用前需看场景。

总1:品牌技术形象。美团以LongCat-2.5强化国产算力与智能体技术品牌:1.6万亿参数、全程国产芯片、未用英伟达GPU;前代LongCat-2.0集群超5万张国产AI芯片并于7月初开源。

1. 产品矩阵:LongCat-2.5主攻编程智能体,LongCat-Image为60亿参数开源文生图且效率优先,CatPaw企业智能体平台7月底上线,官方称覆盖9万员工、3万智能体。

2. 定价与渠道:权重MIT许可,API限时价每百万输入0.27欧元、输出1.07欧元;API兼容OpenAI与Anthropic,可接Claude Code、Codex、Hermes;华为昇腾、摩尔线程、沐曦完成推理适配。

3. 竞争参照:Terminal Bench v2.1得88.4,接近Claude Opus 5的89.1和GPT-5.6 Sol的88.8,低于Deepseek-V4.1-Flash的90.6;DeepSWE 64.6落后Deepseek的74.2;BrowseComp 91.9最高但由美团自有环境运行;v4.0 38.5明显低于Opus 5的53.9。传播宜突出国产芯片、开源、性价比,不宜宣称全面领先。

总1:机会提示。LongCat-2.5以MIT许可证开源,允许商用、修改与再分发,卖家可自托管或二次开发;API兼容OpenAI与Anthropic,可接Claude Code、Codex、Hermes,集成门槛低。

1. 成本机会:限时价每百万输入token约0.27欧元、输出1.07欧元;4比特量化使提示处理吞吐量提升43.3%、响应生成提升42.77%,官方称精度与未压缩模型持平,利于压低推理成本。

2. 增长市场:模型针对国产芯片原生优化,前代集群超5万张国产AI芯片,华为昇腾、摩尔线程、沐曦完成推理适配;CatPaw平台覆盖9万员工、3万智能体,企业智能体需求可跟。

总2:风险与学习点。

1. 风险提示:基准并非全面领先,Terminal Bench v2.1得88.4,低于Deepseek-V4.1-Flash的90.6;DeepSWE 64.6低于Deepseek的74.2和GPT-5.6 Sol的72.7;v4.0仅38.5,明显低于Opus 5的53.9;BrowseComp 91.9由美团自有环境运行,需独立验证。

2. 可学习点:用GitHub真实拉取请求任务、断网并删除Git历史与缓存、由语言模型核查解题过程来防投机,可作为智能体产品评测与训练参考。

总1:生产与部署需求。LongCat-2.5针对国产芯片原生优化,专家权重与激活值采用4比特量化;自托管指南描述194块华为昇腾Atlas A2集群,2块处理图像、64块负责提示摄取、128块负责响应生成,未给出英伟达GPU配置方案。

1. 效率数据:与8比特版本相比,提示处理吞吐量提升43.3%,响应生成吞吐量提升42.77%,官方口径精度与未压缩模型持平,量化有助于降低算力占用。

2. 硬件机会:前代LongCat-2.0集群规模超5万张国产AI芯片,华为昇腾、摩尔线程、沐曦同期完成推理适配;国产芯片厂商可围绕算子适配、通信优化、分布式稳定性形成配套需求。

总2:数字化启示。

1. CatPaw企业智能体平台7月底上线,官方称覆盖9万名员工、搭建3万个智能体,工厂可参考其内部智能体落地路径。

2. 美团2023年起从千卡规模探索国产算力,逐步攻克算子适配、通信优化、分布式稳定性,提示工厂数字化可先从国产硬件适配与量化部署入手;LongCat-Image仅60亿参数、优先效率,也体现轻量实用思路。

总1:行业趋势与新技术。开源权重模型正与国产芯片深度绑定:LongCat-2.5为1.6万亿参数MoE智能体模型,每token仅激活约480亿参数,训练全程国产芯片;图像直接转token与文本同管线,借鉴LongCat-Next,面向视觉编程、设计和三维任务。

1. 训练方法:强化学习用GitHub拉取请求真实软件任务并复现失败案例,断网、删除Git历史与缓存、由语言模型核查解题过程,再把多个教师模型的智能体、推理、对话能力蒸馏进主模型。

2. 部署技术:4比特量化使提示处理吞吐量提升43.3%、响应生成提升42.77%,自托管为194块昇腾Atlas A2分工集群。

总2:客户痛点与解决方案。

1. 痛点:缺英伟达GPU、推理成本高、长链路任务难落地。方案:MIT开源权重可商用修改再分发,API一次最多100万token、单次响应最长128000 token,兼容OpenAI与Anthropic,可接Claude Code、Codex、Hermes,longcat.ai可试用。

2. 风险:基准有强有弱,v4.0仅38.5低于Opus 5的53.9,BrowseComp 91.9为自有环境结果,选型需按场景验证。

总1:平台做法与运营。美团longcat.ai提供对话试用,API兼容OpenAI与Anthropic两种格式,可接入Claude Code、Codex、Hermes等智能体工具;上线期限时促销价每百万输入token约0.27欧元、输出1.07欧元,以低价和兼容性拉新。

1. 开源生态运营:权重以MIT许可证发布,允许商用修改再分发,量化版本放独立代码仓库并提供自托管指南;前代曾以Owl Alpha匿名在OpenRouter运行,可视为先测水温再正式发布的运营手法。

2. 企业平台:CatPaw于7月底上线,官方称覆盖9万员工、3万智能体,显示平台从模型API向企业智能体工作台延伸。

总2:招商与风向。

1. 招商合作:华为昇腾、摩尔线程、沐曦股份等国产芯片厂商完成推理适配,平台可围绕国产算力做软硬件联合;194块昇腾Atlas A2自托管方案给出2块图像、64块提示摄取、128块响应生成的分工。

2. 风向规避:基准并非全面领先,DeepSWE 64.6低于Deepseek的74.2,v4.0 38.5明显低于Opus 5的53.9,BrowseComp 91.9由自有环境运行;宣传应标注测试环境与能力边界,避免过度承诺。

总1:产业新动向。美团LongCat-2.5以1.6万亿参数MoE智能体模型开源,训练全程国产芯片、未用英伟达GPU;前代集群超5万张国产AI芯片。Kimi K3.1、GLM-5.3等国产开源模型已在常见基准基本追平美国头部,针对国产芯片优化的开源权重让相关硬件在海外更具吸引力。

1. 技术路线:每token激活约480亿参数;图像直接转token与文本同管线,借鉴LongCat-Next;4比特量化专家权重与激活值,吞吐提升43.3%和42.77%;强化学习用GitHub PR真实任务,断网删Git历史缓存、语言模型核查解题防投机,并蒸馏多个教师模型能力。

总2:新问题与启示。

1. 评测公平性:BrowseComp 91.9为自有环境最高分,Terminal Bench v2.1 88.4低于Deepseek-V4.1-Flash的90.6,DeepSWE 64.6低于Deepseek的74.2,v4.0 38.5明显低于Opus 5的53.9,说明长链路编程仍有差距,自报环境结果需独立复核。

2. 商业与政策启示:MIT许可允许商用修改再分发,API兼容OpenAI与Anthropic,限时价每百万输入0.27欧元、输出1.07欧元,叠加自托管与CatPaw企业平台,呈现开源权重加API加企业智能体的模式;国产算力从千卡起步攻克算子适配、通信优化、分布式稳定性,可作自主算力生态案例。

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

Summary 1: Core message: Meituan has open-sourced LongCat-2.5, a 1.6-trillion-parameter model designed for long-horizon, multi-step agents, especially coding, and trained entirely on domestic chips without NVIDIA GPUs.

1. Mixture-of-experts architecture activates only about 48 billion parameters per token; the API handles up to 1 million tokens per request and returns responses up to 128,000 tokens, supports text, image, and video input with text output, and offers a toggleable reasoning mode.

2. Images are converted directly into tokens in the same pipeline as text, aiming to improve performance in visual coding, design, and 3D tasks.

Summary 2: Practical use.

1. Available for trials at longcat.ai; the API is compatible with OpenAI and Anthropic and can connect to Claude Code, Codex, and Hermes; weights are under the MIT license, allowing commercial use, modification, and redistribution; limited-time pricing is about €0.27 per million input tokens and €1.07 per million output tokens.

2. The self-hosted quantized version has a separate code repository; the guide uses 194 Huawei Ascend Atlas A2 units—2 for images, 64 for prompt ingestion, and 128 for response generation; 4-bit quantization increases prompt-processing throughput by 43.3% and response-generation throughput by 42.77%, and the company says accuracy matches the uncompressed model.

3. On benchmarks, it scores 88.4 on Terminal Bench v2.1 and improves DeepSWE from 9.7 to 64.6, but only 38.5 on v4.0, below Opus 5's 53.9; use cases should be checked before adoption.

Summary 1: Brand technology positioning. Meituan uses LongCat-2.5 to strengthen its brand around domestic compute and agent technology: 1.6 trillion parameters, trained entirely on domestic chips, no NVIDIA GPUs; the previous LongCat-2.0 cluster used more than 50,000 domestic AI chips and was open-sourced in early July.

1. Product matrix: LongCat-2.5 focuses on coding agents; LongCat-Image is a 6-billion-parameter open-source text-to-image model prioritizing efficiency; the CatPaw enterprise agent platform launched in late July and is said to cover 90,000 employees and 30,000 agents.

2. Pricing and channels: weights are MIT-licensed; limited-time API pricing is €0.27 per million input tokens and €1.07 per million output tokens; the API is compatible with OpenAI and Anthropic and can connect to Claude Code, Codex, and Hermes; Huawei Ascend, Moore Threads, and MetaX have completed inference adaptation.

3. Competitive reference: 88.4 on Terminal Bench v2.1, close to Claude Opus 5's 89.1 and GPT-5.6 Sol's 88.8, but below Deepseek-V4.1-Flash's 90.6; DeepSWE 64.6 trails Deepseek's 74.2; BrowseComp 91.9 is the highest but was run in Meituan's own environment; v4.0 38.5 is clearly below Opus 5's 53.9. Communications should emphasize domestic chips, open source, and cost-effectiveness, not claim overall leadership.

Summary 1: Opportunity signal. LongCat-2.5 is open-sourced under the MIT license, allowing commercial use, modification, and redistribution; sellers can self-host or build on it; API compatibility with OpenAI and Anthropic and support for Claude Code, Codex, and Hermes lower integration barriers.

1. Cost opportunity: limited-time pricing is about €0.27 per million input tokens and €1.07 per million output tokens; 4-bit quantization raises prompt-processing throughput by 43.3% and response-generation throughput by 42.77%, with official accuracy said to match the uncompressed model, helping reduce inference costs.

2. Growth markets: the model is natively optimized for domestic chips; the predecessor cluster exceeded 50,000 domestic AI chips; Huawei Ascend, Moore Threads, and MetaX have completed inference adaptation; CatPaw covers 90,000 employees and 30,000 agents, indicating enterprise-agent demand to follow.

Summary 2: Risks and lessons.

1. Risk note: benchmarks do not show overall leadership. Terminal Bench v2.1 scores 88.4, below Deepseek-V4.1-Flash's 90.6; DeepSWE 64.6 trails Deepseek's 74.2 and GPT-5.6 Sol's 72.7; v4.0 is only 38.5, well below Opus 5's 53.9; BrowseComp 91.9 was run in Meituan's own environment and needs independent verification.

2. Lessons: using real GitHub pull-request tasks, running offline with Git history and caches deleted, and having a language model verify the solution process to prevent gaming can serve as references for agent product evaluation and training.

Summary 1: Production and deployment requirements. LongCat-2.5 is natively optimized for domestic chips, with 4-bit quantization of expert weights and activations; the self-hosting guide describes a 194-unit Huawei Ascend Atlas A2 cluster—2 for image processing, 64 for prompt ingestion, and 128 for response generation—and does not provide an NVIDIA GPU configuration.

1. Efficiency data: compared with the 8-bit version, prompt-processing throughput is up 43.3% and response-generation throughput up 42.77%; officially, accuracy matches the uncompressed model, and quantization helps reduce compute requirements.

2. Hardware opportunity: the predecessor LongCat-2.0 cluster exceeded 50,000 domestic AI chips, and Huawei Ascend, Moore Threads, and MetaX completed inference adaptation at the same time; domestic chip vendors can build supporting demand around operator adaptation, communication optimization, and distributed stability.

Summary 2: Digitalization lessons.

1. The CatPaw enterprise agent platform launched in late July and is said to cover 90,000 employees and 30,000 agents; factories can reference its internal agent deployment path.

2. Since 2023, Meituan has explored domestic compute from thousand-card scale, gradually overcoming operator adaptation, communication optimization, and distributed stability; this suggests factories can start digitalization with domestic hardware adaptation and quantized deployment; LongCat-Image, with only 6 billion parameters and efficiency first, also reflects a lightweight, practical approach.

Summary 1: Industry trends and new technology. Open-weight models are becoming tightly coupled with domestic chips: LongCat-2.5 is a 1.6-trillion-parameter MoE agent model activating about 48 billion parameters per token and trained entirely on domestic chips; images are converted directly to tokens in the same pipeline as text, drawing on LongCat-Next and targeting visual coding, design, and 3D tasks.

1. Training method: reinforcement learning uses real software tasks from GitHub pull requests and reproduces failure cases; it runs offline, deletes Git history and caches, and uses a language model to verify the solution process, then distills agentic, reasoning, and dialogue capabilities from multiple teacher models into the main model.

2. Deployment technology: 4-bit quantization raises prompt-processing throughput by 43.3% and response-generation throughput by 42.77%; self-hosting uses a divided 194-unit Ascend Atlas A2 cluster.

Summary 2: Customer pain points and solutions.

1. Pain points: lack of NVIDIA GPUs, high inference costs, and difficulty deploying long-horizon tasks. Solution: MIT-licensed open weights allow commercial use, modification, and redistribution; the API handles up to 1 million tokens per request and responses up to 128,000 tokens, is compatible with OpenAI and Anthropic, connects to Claude Code, Codex, and Hermes, and is available for trial at longcat.ai.

2. Risk: benchmark results are mixed; v4.0 scores only 38.5 versus Opus 5's 53.9, and BrowseComp 91.9 is a result from Meituan's own environment; selection should be validated by use case.

Summary 1: Platform practices and operations. Meituan's longcat.ai offers chat-based trials; the API is compatible with both OpenAI and Anthropic formats and can connect to agent tools such as Claude Code, Codex, and Hermes; introductory limited-time pricing is about €0.27 per million input tokens and €1.07 per million output tokens, using low price and compatibility to attract users.

1. Open-source ecosystem operations: weights are released under the MIT license, allowing commercial use, modification, and redistribution; the quantized version is placed in a separate code repository with a self-hosting guide; the predecessor ran anonymously as Owl Alpha on OpenRouter, which can be seen as testing the waters before an official launch.

2. Enterprise platform: CatPaw launched in late July and is said to cover 90,000 employees and 30,000 agents, showing the platform extending from model APIs to an enterprise agent workbench.

Summary 2: Partnerships and trend caution.

1. Partnership opportunities: domestic chip vendors including Huawei Ascend, Moore Threads, and MetaX have completed inference adaptation, so the platform can pursue software-hardware collaboration around domestic compute; the 194-unit Ascend Atlas A2 self-hosting plan assigns 2 units to images, 64 to prompt ingestion, and 128 to response generation.

2. Trend caution: benchmarks do not show overall leadership; DeepSWE 64.6 trails Deepseek's 74.2; v4.0 38.5 is clearly below Opus 5's 53.9; BrowseComp 91.9 was run in its own environment; promotional materials should specify test environments and capability boundaries to avoid overpromising.

Summary 1: New industry developments. Meituan has open-sourced LongCat-2.5, a 1.6-trillion-parameter MoE agent model trained entirely on domestic chips without NVIDIA GPUs; the predecessor cluster exceeded 50,000 domestic AI chips. Domestic open models such as Kimi K3.1 and GLM-5.3 have essentially caught up with leading U.S. models on common benchmarks, and open weights optimized for domestic chips make related hardware more attractive overseas.

1. Technical route: about 48 billion parameters activated per token; images are converted directly to tokens in the same pipeline as text, drawing on LongCat-Next; 4-bit quantization of expert weights and activations raises throughput by 43.3% and 42.77%; reinforcement learning uses real GitHub PR tasks, runs offline with Git history and caches deleted, uses language-model verification to prevent gaming, and distills capabilities from multiple teacher models.

Summary 2: New questions and implications.

1. Evaluation fairness: BrowseComp 91.9 is the highest score but was run in Meituan's own environment; Terminal Bench v2.1 88.4 is below Deepseek-V4.1-Flash's 90.6; DeepSWE 64.6 is below Deepseek's 74.2; v4.0 38.5 is clearly below Opus 5's 53.9, indicating that long-horizon coding still has gaps, and self-reported environment results need independent review.

2. Commercial and policy implications: the MIT license allows commercial use, modification, and redistribution; the API is compatible with OpenAI and Anthropic; limited-time pricing is €0.27 per million input tokens and €1.07 per million output tokens; combined with self-hosting and the CatPaw enterprise platform, this presents an open-weights + API + enterprise-agent model; domestic compute started from thousand-card scale and overcame operator adaptation, communication optimization, and distributed stability, offering a case study for an autonomous compute ecosystem.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

美团LongCat团队在官方发布一周后正式公开LongCat-2.5的模型权重。该模型总参数规模1.6万亿,专为自主完成长链路、多步骤任务的智能体设计,尤其面向编程场景,训练全程基于中国国产芯片完成,未使用英伟达GPU。

LongCat-2.5采用混合专家架构,每个token仅激活全部参数中的约480亿个,通过路由调用部分专用子网络,而非运行整个模型。经API接入时,模型可一次处理最多100万个token,单次响应最长128000个token。它能理解文本、图像和视频,以文本形式输出,并允许用户自行开关推理模式。

与事后拼接独立图像编码器的常见做法不同,LongCat-2.5直接将图像转换为token,与文本在同一管线中处理。这一设计借鉴了研究模型LongCat-Next,目标是提升视觉编程、设计及三维任务的表现。在强化学习阶段,团队使用来自GitHub拉取请求的真实软件任务,并复现了智能体的失败案例。为确保模型在训练中真正解决问题而非投机取巧,开发人员在训练环境中断开互联网访问,删除Git历史记录与缓存文件,再由语言模型核查解题过程。团队还将多个专用教师模型在智能体、推理和对话方面的能力蒸馏进LongCat-2.5。

针对国产芯片的原生特性,美团对专家权重与激活值采用4比特量化。负载测试显示,与8比特版本相比,提示处理吞吐量提升43.3%,响应生成吞吐量提升42.77%,官方口径为精度与未压缩模型持平。该量化版本在独立代码仓库中提供,自托管指南描述了一个由194块华为昇腾Atlas A2系列加速卡组成的集群,其中2块处理图像、64块负责提示摄取、128块负责响应生成,指南未给出英伟达GPU的配置方案。

基准测试方面,LongCat-2.5在Terminal Bench v2.1上得88.4分,Claude Opus 5为89.1分,GPT-5.6 Sol为88.8分,Deepseek-V4.1-Flash以90.6分居首。在测试模型修复真实代码仓库缺陷能力的DeepSWE上,LongCat-2.5从9.7分提升至64.6分,仅高于Claude Opus 4.8,Deepseek-V4.1-Flash以74.2分领先,GPT-5.6 Sol为72.7分。难度更高的Terminal Bench v4.0中,该模型得38.5分,略高于GPT-5.6 Sol的37.3分,但明显低于Opus 5的53.9分。网页搜索基准BrowseComp上,LongCat-2.5得91.9分,为表格中最高,不过该项测试由美团在自有环境中运行。

权重以MIT许可证发布,允许商用、修改与再分发。用户可在longcat.ai的对话界面试用,API同时兼容OpenAI与Anthropic两种格式,可接入Claude Code、Codex、Hermes等智能体工具。上线期间美团提供限时促销价,约合每百万输入token 0.27欧元、每百万输出token 1.07欧元。

LongCat-2.5的前代LongCat-2.0于今年6月底完成训练,集群规模超过5万张国产AI芯片,并在7月初开源,华为昇腾、摩尔线程、沐曦股份等国产芯片厂商同期宣布完成推理适配。更早之前,该模型曾以“Owl Alpha”的匿名身份在OpenRouter上运行。公开资料显示,美团LongCat团队对国产算力的探索始于2023年,从千卡规模起步,逐步攻克算子适配、通信优化、分布式稳定性等问题。基于LongCat 2.0构建的全场景智能体平台CatPaw于7月底上线,官方信息称其已覆盖9万名员工、搭建3万个智能体。美团同期还在推进图像生成模型,此前发布的LongCat-Image是一款仅60亿参数的开源文生图模型,优先考虑效率而非模型规模。

在开源权重阵营中,Kimi K3.1、GLM-5.3等国产模型已在常见基准上基本追平美国头部模型。从设计之初便针对国产芯片优化的开源权重模型,也让这些硬件在海外市场更具吸引力。

本文首发于 亿邦动力 官方网站

【版权提示】亿邦动力倡导尊重与保护知识产权。未经许可,任何人不得复制、转载、或以其他方式使用本网站的内容。如发现本站文章存在版权问题,烦请提供版权疑问、身份证明、版权证明、联系方式等发邮件至run@ebrun.com,我们将及时沟通与处理。

已点赞

+1

{{ praise_text }}

好文章,值得被更多人看到

FAQ回顾

美团LongCat-2.5是什么模型?

美团LongCat-2.5是美团LongCat团队开源的1.6万亿参数智能体模型,专为自主完成长链路、多步骤任务设计,尤其面向编程场景。它采用混合专家架构,每个token仅激活约480亿参数,支持文本、图像和视频输入并以文本输出,用户可自行开关推理模式。训练全程基于中国国产芯片完成,未使用英伟达GPU,权重以MIT许可证发布。

LongCat-2.5如何适配国产AI芯片?

LongCat-2.5针对国产芯片原生特性,对专家权重与激活值采用4比特量化。自托管指南描述了一个由194块华为昇腾Atlas A2系列加速卡组成的集群,其中2块处理图像、64块负责提示摄取、128块负责响应生成。负载测试显示,与8比特版本相比,提示处理吞吐量提升43.3%,响应生成吞吐量提升42.77%,官方口径为精度与未压缩模型持平。前代LongCat-2.0使用超5万张国产AI芯片,并解决算子适配、通信优化和分布式稳定性问题。

LongCat-2.5在编程智能体任务上的基准表现如何?

在Terminal Bench v2.1上,LongCat-2.5得88.4分,Claude Opus 5为89.1分,GPT-5.6 Sol为88.8分,Deepseek-V4.1-Flash以90.6分居首。在测试修复真实代码仓库缺陷能力的DeepSWE上,LongCat-2.5从9.7分提升至64.6分,仅高于Claude Opus 4.8,低于Deepseek-V4.1-Flash的74.2分和GPT-5.6 Sol的72.7分。Terminal Bench v4.0中得38.5分,略高于GPT-5.6 Sol的37.3分,但低于Opus 5的53.9分。

企业如何试用或接入LongCat-2.5?

用户可在longcat.ai的对话界面试用LongCat-2.5。API同时兼容OpenAI与Anthropic两种格式,可接入Claude Code、Codex、Hermes等智能体工具。上线期间美团提供限时促销价,约每百万输入token 0.27欧元、每百万输出token 1.07欧元。权重以MIT许可证发布,允许商用、修改与再分发;自托管可参考独立代码仓库中的量化版本和指南,指南使用华为昇腾Atlas A2系列加速卡集群。

美团LongCat系列在国产算力和智能体平台上有哪些布局?

美团LongCat团队的国产算力探索始于2023年,从千卡规模起步,逐步攻克算子适配、通信优化、分布式稳定性等问题。LongCat-2.0于今年6月底完成训练,集群规模超过5万张国产AI芯片,7月初开源,华为昇腾、摩尔线程、沐曦股份等宣布完成推理适配。基于LongCat 2.0的全场景智能体平台CatPaw于7月底上线,已覆盖9万名员工、搭建3万个智能体。LongCat-Image是60亿参数开源文生图模型。

广告

下一篇

胖东来洗发水事件背后:商业责任边界该如何厘清?

让制造者承担制造的责任,让品牌方管好品牌的故事,让零售商守住真诚与兜底的底线,这才是用理性的眼光去看清商业逻辑。