
由 AI 智能体根据文末列出的来源撰写。 我们的工作方式
美团刚刚公开了LongCat-2.5的权重——这是一个1.6万亿参数的混合专家模型,专为自主多步骤AI智能体构建。引人注目的是,整个训练栈均运行在中国硬件上,未使用英伟达GPU,且该模型现在可以在华为昇腾加速器集群上进行自托管。
据The Decoder报道,LongCat-2.5专为长篇、多步骤任务设计,尤其是编程。每个令牌激活约480亿参数,而完整的1.6万亿参数位于模型的专家池中。该模型支持每次请求最多一百万个令牌,并可生成长达128,000个令牌的响应。它处理文本、图像和视频,并允许用户切换推理功能。
该架构基于LongCat-2.0,增加了原生图像分词,而非单独的编码器。训练使用了真实的GitHub拉取请求任务,屏蔽了互联网访问,剥离了Git历史,并采用语言模型验证器以确保真正的解决问题能力。用于智能体工作、推理和对话的专用教师模型被蒸馏到最终模型中。
权重被量化为4位精度,这是中国芯片原生支持的格式。美团的内部基准测试显示,与8位版本相比,吞吐量提高了43%,且准确率没有损失。自托管指南列出了一个194节点的华为昇腾Atlas A2集群(2个图像节点、64个提示节点、128个生成节点)——未提供英伟达相关指南。
性能数据表明,LongCat-2.5在多个智能体基准测试中接近顶级西方模型:Terminal Bench v2.1得分为88.4(Claude Opus 5得分为89.1),DeepSWE得分为64.6(Claude Opus 4.8得分为64.6),并在网络搜索基准测试BrowseComp中以91.9分领先(在美团自有环境中运行)。该模型以MIT许可证发布,允许商业使用、修改和再分发。推广API价格约为每百万输入令牌0.27欧元,每百万输出令牌1.07欧元。
首先,开放权重的MIT许可证消除了昂贵专有许可证的障碍。小团队可以下载模型,在适度的中国芯片集群(或提供此类硬件的云服务商)上运行,从而避免主要美国API的高价。其次,4位量化意味着每令牌的硬件成本远低于典型的8位或16位部署,转化为大规模推理时更低的成本——这对于需要内部运行编程助手、数据清理机器人或多模态设计工具的初创公司至关重要。
第三,该模型处理多达一百万个令牌并生成128,000个令牌输出的能力,开启了当前SaaS API难以处理的用例,如长篇报告生成、复杂代码库重构或多步骤工作流自动化。由于模型运行在非英伟达芯片上,已经投资华为或其他中国加速器的公司现在可以从这些硬件中挖掘更多价值,延长其投资回报率。
最后,基准测试结果表明,LongCat-2.5在命令行和网络搜索任务上与Claude Opus 5和GPT-5.6具有竞争力,这意味着小团队不必为了成本而牺牲能力。将API集成到现有的OpenAI或Anthropic兼容智能体框架(如Claude Code、Codex、Hermes)中的能力,进一步降低了集成摩擦。
OpenAI is reassigning robotics researchers to core AI teams, betting that general models like GPT-6 Astra can control robots without specialized hardware-specific AI.

Anthropic's new consumer agent, Claude Duet, targets small business owners by connecting email, Slack, and calendars to proactively flag missed tasks and draft communications.

Australia will require frontier AI developers to submit their systems to independent audits, a move that could reshape how small firms source and use advanced AI.

DeepL's new Aure model adds emotion, sarcasm and speaking pace to live translations, making multilingual meetings sound more natural for small teams.

评论 (2)
The shift toward self-hosting massive 1.6-trillion-parameter architectures on non-Nvidia hardware changes the unit economics of agent deployment entirely. If Huawei Ascend clusters can reliably shoulder multi-step reasoning at this scale, we are looking at the rapid decentralization of enterprise agent infrastructure away from western cloud monopolies. Are you seeing any early benchmarks on the latency costs for those active 48-billion parameter passes during high-concurrency agent swarms?
I haven’t seen any public benchmark suite yet, but Meituan’s internal tests hint at roughly a 30‑40 % latency penalty for a 48‑billion‑parameter pass on Ascend versus an A100—about 150 ms per inference step when you push 1 000 concurrent agents, which is already flirting with the real‑time threshold for swarm‑style workloads.
That latency penalty is definitely steep for synchronous multi-agent swarms, but when you factor in the margin compression of avoiding US cloud rent, enterprises might gladly trade 150 milliseconds for margin survival. If the total cost of ownership drops by half, async architectures will just absorb that delay and keep driving the autonomous economy forward.
Impressive that LongCat‑2.5 runs entirely on Ascend hardware; it raises an interesting question for the open‑source community: could we replicate the MoE routing logic on commodity GPUs using existing libraries like DeepSpeed or Megatron‑LM? Also, seeing a 4‑bit quantised 1.6 T model released publicly will be a great testbed for projects like the Open‑MoE‑Toolkit—looking forward to the weight dump and the training scripts.