一个 1.6 万亿参数的模型，每处理一个 token 只唤醒其中约 48B，也就是 3%。美团在 11 月 1 日把 LongCat-2.5 正式版放了出来，输入定价 0.3 美元每百万 token，缓存命中只要 0.006 美元。这个价格在同级模型里直接砍到了脚踝，而且它还带着 100 万 token 的原生上下文和视觉理解能力。外卖公司做模型已经不是新闻，但这次美团瞄准的不是聊天，是 Agent 干活：终端命令、浏览器操作、GUI 界面、表格和设计工具，全是长周期任务里最磨人的场景。这篇文章把发布内容、稀疏激活的成本逻辑、接入方式和切换判断一次讲清楚。
一周两个版本：Preview 到正式版改了什么
LongCat-2.5 不是突然冒出来的。9 月 25 日美团先放出了 Preview 版，当时只开放 API 调用，没有公布完整技术报告，权重也没有开放。五周之后的 11 月 1 日，正式版上线，定价、上下文、输出上限这些完整规格随之公开。
核心规格先摆出来：MoE 架构，总参数 1.6T，每个 token 激活约 48B；原生支持多模态输入，能直接看图做跨模态问答和视觉推理；上下文窗口 100 万 token 输入，输出上限 13.1 万 token。定位非常明确，就是给长周期 Agentic 任务用的，官方点名适配的场景包括终端、浏览器、GUI 操作、表格处理和设计工具。
值得留意的是适配名单。LongCat-2.5 深度适配了 Claude Code、Hermes、OpenClaw、OpenCode 和 Kilo Code 这几个主流编码 Agent。这份名单说明美团没打算让模型只活在自家聊天框里，而是直接瞄准开发者已经在用的工具链。模型好不好用是一回事，能不能装进现有工作流是另一回事，美团显然想在第二条路上抢跑。
和前代对比能看出路线变化。2025 年 9 月开源的 LongCat-Flash-Chat 是 560B 总参、平均激活约 27B，主打推理速度。一年多时间总参数翻了接近三倍，激活比例反而压得更低，还补上了视觉能力。美团的路线很清晰：不靠堆激活参数换质量，而是用更大的专家池子加更狠的稀疏化，把成本压下来。
稀疏激活的账：1.6T 为什么只按 48B 收钱
MoE 这个架构已经讲了几年，但 LongCat-2.5 把稀疏度推到了一个新比例。1.6T 总参数、48B 激活，意味着每次前向计算只动用 3% 的参数，剩下 97% 待在原地，不产生计算开销。
这个比例直接决定推理成本的结构。模型推理的算力消耗大致和激活参数量成正比，而不是和总参数量成正比。一个 48B 激活的 MoE 模型，单 token 的计算量约等于一个 48B 稠密模型，但质量上限由整个 1.6T 的专家池决定。用通俗的话说，你付的是小模型的电费，用的是大模型的脑子。
定价印证了这个逻辑。LongCat-2.5 输入 0.3 美元每百万 token，缓存命中的前缀只要 0.006 美元，输出 1.2 美元。缓存输入价和全价输入差了 50 倍，这个价差对 Agent 场景极其重要：Agent 干活时会反复携带大段历史上下文，系统提示词和工具定义每次都一样，缓存命中率天然就高。一个跑了一整天的编码 Agent，输入 token 的大头都能按缓存价结算，实际成本和标价完全是两个概念。
这也是为什么美团敢把 100 万上下文开放出来。长上下文最大的成本不是显存，而是每次调用都要重新处理前缀。稀疏激活加前缀缓存，把这两项成本都压进了可接受的区间，长上下文才从实验室指标变成能用得起的产品参数。
还有一个容易被忽略的细节：稀疏度越高，对专家路由的要求越苛刻。1.6T 参数里每次只挑 48B 出来干活，挑错了专家，输出质量立刻掉档。美团敢把激活比例压到 3%，说明路由层的训练已经过了一道硬门槛，这不是堆卡就能解决的问题，是实打实的工程积累。
1M 上下文不是噱头：长周期 Agent 的真实瓶颈
聊 100 万上下文之前，先看看 Agent 任务到底怎么消耗 token。一个编码 Agent 处理真实需求时，上下文里要塞的东西包括：系统提示词、工具定义、代码库检索结果、历史对话、上一步的工具输出、错误堆栈。跑一个中等复杂度的任务，几十轮交互下来轻松突破 20 万 token。
上下文不够时会发生什么？Agent 开始遗忘。它忘了文件开头定义的接口，忘了三轮之前用户提的约束，甚至忘了自己刚才改过哪个文件。外在表现就是代码前后矛盾、重复修改、把已经修好的 bug 又改回去。市面上的应对方案是压缩和摘要，但摘要必然丢信息，丢了什么、丢多少，完全不可控。
100 万 token 的原生上下文意味着大多数真实项目可以整个塞进去。一个中型代码库的核心源码、完整的会话历史、几十份参考文档，全部留在上下文里，不需要中途裁剪。13.1 万的输出上限也够让模型一次性吐出完整的大文件，不用分段生成再人工拼接。
当然，上下文窗口和有效利用是两回事。模型在超长输入里定位关键信息的能力会衰减，这要靠自己的任务集实测验证。但至少在容量层面，LongCat-2.5 把天花板抬到了当前已知模型的前列，容量瓶颈不再是推脱任务失败的理由。
原生多模态进来之后，工作流少了哪几步
LongCat-2.5 的视觉能力是原生的，不是外挂一个视觉编码器再拼接。区别体现在延迟和理解深度上：原生多模态在训练阶段就联合学习图文，对界面截图、架构图、表格照片的理解更连贯，不需要跨模型翻译。
对 Agent 工作流来说，这个能力补上了最尴尬的一环。以前 Agent 操作 GUI 软件时，需要先截图，再用单独的视觉模型描述画面，把描述文本喂给语言模型。三步走下来，信息损耗加延迟，错误率在每一步都在累积。原生多模态让模型直接看截图做决策，省掉了中间的文本转述，错误链条短了一大截。
落到具体场景就很实际。让 Agent 处理一份扫描的表格，以前是 OCR 加文本解析加人工核对，现在是截图加一句指令。让 Agent 调试前端页面，以前是费劲描述渲染异常，现在是直接把页面截图给它看。官方列出的适配场景里，GUI 操作和设计工具都在名单上，说明美团自己就把视觉理解当成 Agent 闭环的一部分，而不是发布会上凑数的附加功能。
多模态还改变了一件事：Agent 的感知范围。只能读文本的 Agent 只能活在终端和编辑器里，能看图的 Agent 才能碰真实世界的软件。从这个角度看，原生视觉不是加分项，是长周期 Agent 的入场券。
接入实测：官方 API 与现有 Agent 框架
目前 LongCat-2.5 通过美团官方 API 提供服务，调用方式是业界通用的 OpenAI 兼容格式，基础的文本对话用几行代码就能跑起来：
from openai import OpenAI
import os
client = OpenAI(
api_key=os.environ["LONGCAT_API_KEY"],
base_url=os.environ["LONGCAT_BASE_URL"],
)
resp = client.chat.completions.create(
model="longcat-2.5",
messages=[
{"role": "system", "content": "你是一个资深后端工程师"},
{"role": "user", "content": "把这个接口重构成异步版本"},
],
)
print(resp.choices[0].message.content)
具体的基础地址和鉴权细节以美团官方文档为准，注册后按指引创建密钥即可。调用结构和 OpenAI SDK 完全一致，已有代码把客户端初始化换掉就能跑通，迁移成本基本为零。
更有价值的是 Agent 框架侧的支持。Claude Code、OpenClaw、OpenCode 这些工具接入新模型通常只需要改两个配置：模型名和接口地址。官方明确列出适配名单，意味着工具调用层面的兼容性已经做过验证，不需要自己处理 function calling 的格式差异。踩过工具调用格式坑的人都知道，这一项能省掉好几天的调试时间。
对国内团队还有一个隐性优势：API 走国内基础设施，网络层面的不确定性少了一截。延迟和可用性这两件事，只有跑过生产环境的人才知道有多重要，聊天 demo 里感觉不出来，Agent 连续跑八小时就全暴露了。
摆数据：LongCat-2.5 与前代的对比
把核实过的数据放在一起看，比宣传稿更能说明问题：
两条信息值得单独拎出来。一条是激活参数从 27B 涨到 48B，说明美团没有为了账面参数牺牲单 token 质量，激活量的提升会实实在在反映在复杂推理能力上。另一条是获取方式从开源权重转向官方 API，对想自部署的团队是个约束，对只想调 API 的团队反而省了运维和显卡账单，两种需求不可能同时满足，美团这次选了后者。
从行业位置看，LongCat-2.5 被拿来和 GPT-6 Astra、Claude Opus 4.8、Kimi K3 这一档模型做对比，能进入这个对比序列本身，说明国产模型在 Agent 赛道上已经不是陪跑角色。更值得注意的是对比维度变了：以前比的是跑分和榜单排名，现在比的是激活效率、缓存价差和上下文容量，全是影响实际账单和任务成功率的硬指标。跑分高但用不起的模型，在 Agent 时代没有位置。
三种团队的切换判断
值不值得把现有工作流迁到 LongCat-2.5，取决于你在做什么。
跑长周期 Agent 任务的团队，比如用编码 Agent 处理真实需求、让 Agent 操作内部系统的，值得立刻试。100 万上下文加缓存定价，是为这类场景量身定做的组合，成本结构优势在重型任务里会被放大，跑一周账单就能看出差距。
重度依赖多模态的团队同样值得关注。Agent 需要看截图、读表格、理解界面的，原生视觉能力能把三步流程压成一步，错误率和延迟同时下降。特别是做自动化测试、RPA 和数据录入的团队，GUI 理解能力直接决定了 Agent 能接多少活。
已经深度绑定现有供应商、任务以短对话为主的团队，可以先观望不着急。模型能力终归要靠自己的任务集实测，把日常 prompt 抽一批出来两边跑，对比质量和实际账单，自己的数据比任何第三方评测都有说服力。
LongCat-2.5 的出现说明一件事：Agent 时代的模型竞争，参数规模不再是唯一变量，稀疏度、缓存定价和上下文容量组成的成本账本，正在变成新的主战场。谁能把干活的成本压到最低，谁就能装进最多的工作流。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A model of 1.6 trillion parameters, of which each token is processed, only about 48B, or 3%. The group released the LongCat-2.5 official version on 1 November and entered a price of $0.3 per million token, cache hit for $0.006. This price cuts directly to the ankles in a peer model, and it also carries the original context and visual understanding of 1 million tokens. The modeling company is no longer news, but this time it's not about chat, it's about Agent working: terminal commands, browser operations, GUI interfaces, tables and design tools, all of which are the most grinding scenes of a long-cycle mission. The article explains the content of publish, the cost logic of diluted activation, the mode of access and the rotation of judgement.
Two editions a week: Preview to what?
LongCat-2.5 isn't suddenly coming out. On September 25th, the delegation first released the Preview version, which was open only for API calls, did not publish a full technical report and did not open the weights. Five weeks later, on 1 November, the full specifications of the official version, pricing, context, output caps, were then made public.
Core specification first: MoE architecture, total parameter 1.6T, each token activates approximately 48B; original supports multimodular input, allowing for cross-modular questions and answers and visual reasoning; context window1 million token input, output cap 131,000 token. Positioning is very clear, for long-cycle Agenic missions, and the official roll-calls include terminals, browsers, GUI operations, table processing and design tools.
It is worth noting that the appliance list. LongCat-2.5 Depth fits the mainstream codes Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code. This list indicates that the Corps does not intend to keep the model in its own chat box, but to target the tool chain that the developers are already using. The usefulness of the model is one thing. Whether or not it can be loaded into the existing workflow is another, and it is clear that the group wants to run on the second road.
A comparison with previous generations reveals a change of course. LongCat-Flash-Chat from September 2025 is 560B total gin, an average activation of about 27B, and the main reason speed. The total parameter of more than one year has nearly tripled, and the activation rate has been even lower, adding to the visual ability. The mission's route is clear: instead of using stacking of activation parameters for mass, it is using a larger pool of experts to add a sharper thinning to reduce costs.
Rare-activated account: 1.6T Why only 48B to collect money?
MoE has been talking about this structure for several years, but LongCat-2.5 pushs the thinness to a new scale. 1.6T Total, 48B activated means that only 3% of the parameter is used for each forward calculation, leaving 97% in place and no calculation costs.
This ratio directly determines the structure of the reasoning costs. The computational consumption of model reasoning is roughly proportional to the activation of parameter count, rather than to total parameter count. A 48B activated MoE model, the single token calculation is approximately equal to a 48B dense model, but the maximum mass is determined by the entire 1.6T pool of experts. In plain language, you pay for small models, with LLM brain.
Pricing confirms this logic. LongCat-2.5 Enter $0.3 per million token, cache hit prefix only $0.006, output $1.2. The difference between the cache input price and the full price input is 50 times greater, and this difference is extremely important for the Agent scenario: Agent works with a large history of context, the system prompt is the same as the tool definition each time, and the cache hit rate is natural. An all-day coding, Agent, entered token's head can be settled at cache prices, and the actual costs and bid prices are completely two concepts.
And that's why the American group dared to open 1 million contexts out. The largest cost of long context was not VRAM, but to reprocess prefixes every call. Rarely activated and prefixed cache, both of which were placed in acceptable compartments, was converted from laboratory indicators to affordable product parameters.
There is also an easy-to-neglected detail: the more thinness, the more demanding the experts are. 1.6T parameters only select 48B each time to work, pick the wrong expert, and output quality drops immediately. And the group is willing to push the activation rate to 3%, which means that the routing level has passed a hard threshold. It's not a pile of cards, it's a solid engineering accumulation.
1M context is not the real bottleneck of long cycle Agent
Before we talk about 1 million context, let's see how the Agent mission consumes token. A code Agent handles real needs, forts in context include: system prompt, tool definition, code library search results, historical conversation, tool output in the last step, error stacking. Run a medium-complicated task, couple dozen rounds of interaction down easy token.
What happens when context isn't enough? Agent started forgetting. It had forgotten the interface defined at the beginning of the document, the constraints that had been imposed by the user three times earlier, or even the document it had just changed. The external expression is that the code is inconsistent, re-engineered, and the bugs that have been repaired are returned. The response on the market was compression and summary, but summary must have lost information, lost what, how much, and was completely uncontrollable.
One million token's original context means that most real projects can be plugged in. The core source code, full session history, dozens of reference documents of a medium code library, all of which remain in context without mid-cuts. The output ceiling of 13.1 million is also sufficient to allow the model to throw out the full file at once, without creating a sequence to re-assemble manually.
Of course, context window and its effective use are different. The model ' s ability to locate key information in a super-long input is reduced, which is verified by its own task collection. But at least at the level of capacity, LongCat-2.5 lifts the ceiling to the forefront of currently known models and capacity bottlenecks no longer justify the failure of the task.
How many steps have workflow been missing since he came in?
LongCat-2.5's visual capability is original, not an external visual encoder remix. The difference is reflected in the depth of delay and understanding: native multi-modulars are combined to learn text at the training stage, and there is a more consistent understanding of interfaces, architectures and table photographs, which do not require translation across models.
For Agent workflow, this ability adds to the most embarrassing link. Previously, Agent used to operate GUI software to screen first, then use a separate visual model to describe the image and feed the description to the language model. Three steps down, information losses and delays, and error rates are accumulating at each step. Prototypes allow the model to take a direct look at the screenshot for decision, saves the middle text transliteration and shorts the wrong chain.
It's practical to get to the concrete scene. Allows Agent to process a scanned form, which was used by OCR with text resolution and manual checking, which is now a screenshot with a command. Let Agent debug the front-end page, which used to be hard to describe the rendering anomaly, now shows it directly. In the officially listed fit scenario, the GUI operation and design tools are on the list, indicating that the Corps itself sees visual understanding as part of the Agent closed ring, rather than as an additional function to be counted at publish.
The multiplicity also changed one thing: the sense range of Agent. Agent, which can only read the text, can only touch the software of the real world by living in terminals and editors, with graphics. From this point of view, native vision is not a subdivision, but a ticket to the long cycle Agent.
Access information: Official API and existing Agent framework
LongCat-2.5 is currently served through the Mission's official API, which is used in industry's Open AI compatible format, and the underlying text conversations can run in several lines:
from openai import OpenAI
import os
client = OpenAI(
api_key=os.environ["LONGCAT_API_KEY"],
base_url=os.environ["LONGCAT_BASE_URL"],
)
resp = client.chat.completions.create(
model="longcat-2.5",
messages=[
"You're a senior back-end engineer,"
== sync, corrected by elderman == @elder man
],
)
print(resp.choices[0].message.content)
The specific base address and authentication details are based on the official United States Mission documents, and registration is followed by the creation of a key in accordance with the guidelines. The call structure and Open AI SDK are fully consistent, with the existing code replacing the client's initialization and running, with almost zero migration costs.
More valuable is the Agent framework side support. Claude Code, OpenClaw, OpenCode, and OpenCode usually need only two configurations to access the new model: model name and interface address. The official clear listing of suitability means that compatibility at the level of tool call has been tested and does not have to deal with the differences in the format of the fight calling. Those who stepped on the tool call pit are aware that this one will save days of debugging.
There is also a hidden advantage for national teams: API takes the domestic infrastructure, and there is less uncertainty at the network level. The two things about delay and usability are that only people who run past the production environment know how important it is to talk and demo can't feel it, and Agent runs eight hours in a row.
Data set: LongCat-2.5 against the previous generation
The fact that the verified data are viewed together is more illustrative than the promotional text:
Two messages deserve to be carried out separately. One is to activate the parameter from 27B to 48B, which means that the United States does not sacrifice the token mass for the book parameter, and that the increase in the amount of activation is actually reflected in complex reasoning capabilities. The other is that the way to access is to move from the open source weight to the official API, which is a constraint on the team that wants to deploy itself. Instead, the team that wants to move API saves peacekeeping and GPU bills, two demands that cannot be met at the same time.
LongCat-2.5 is compared to GPT-6 Astra, Claude Opus 4.8, Kimi K3, from industry location, and can enter the comparison itself, suggesting that the national production model is no longer a running role on Agent. More notably, the contrast dimension has changed: previously, running scores and rankings were compared to activation efficiency, cache price differentials and context capacity, all hard indicators affecting actual bills and tasks success rate. High-scoring but unserviceable models have no place in the Age of Agent.
Three team-to-tasks.
It's not worth moving the existing workflow to LongCat-2.5 depending on what you're doing.
Teams that run long-cycle Agent missions, such as using the code Agent to process real needs and allow Agent to operate internal systems, are worth trying immediately. One million context plus cache are priced for a tailor-made combination of such scenes, and the cost structure advantages are magnified in heavy missions, and a one-week billing shows the difference.
The heavy reliance on multi-modular teams also deserves attention. Agent needs to look at the screenshot, read the table, understand the interface, and is able to press the three-step process into one step, with the error rate and delay decreasing simultaneously. Especially the team that did automation tests, the RPA and the data entry, the GUI's ability to understand directly determines how much Agent can do.
Teams with existing suppliers and mandates focused on short dialogue can wait and wait. The model capacity is ultimately based on its own tasks, extracting routine prompts from one another, comparing quality and actual bills, and its own data is more convincing than any third-party assessment.
The emergence of LongCat-2.5 illustrates one thing: the Agent-era model competition, the size of the parameters is no longer the only variable, and the cost book consisting of thinness, cache pricing and context capacity is becoming the new main battlefield. Anyone who can minimize the cost of doing his work can load the most workflow.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.