智东西（公众号：zhidxcom）
作者 | 杨京丽
编辑 | 李水青
智东西11月2日报道，今日，美团LongCat团队发布面向长程智能体任务的原生多模态模型LongCat-2.5。
该模型总参数量为1.6万亿，每个Token激活约480亿参数，在LongCat-2.0的基础上，重点强化原生视觉理解与智能体编程能力。
在美团公布的自动化任务基准AutomationBench、通用Agent任务基准LongCatClawBench等几项基准测评中，LongCat-2.5均领先Claude-Opus-5、GPT-5.6-Sol等模型；在终端操作基准Terminal Bench v4.0上，LongCat-2.5的得分为38.5，超过GPT-5.6-Sol，但仍落后于Claude-Opus-5。
LongCat-2.5的长序列训练中引入更多长程轨迹数据，并通过大规模长程强化学习，提升模型在复杂代码库开发、完整项目构建等任务中的端到端工程能力。同时，LongCat-2.5将视觉信号转化为多层级离散Token，让视觉与文本输入由同一主干模型统一处理，进一步支持视觉编程（Visual Coding）、创意设计和3D内容生成。
面对国产加速器单卡不足80GB的显存限制，LongCat-2.5围绕超节点重排FSDP、EP、CP并行度，叠加混合词表并行、MoonEP动态专家复制和双重分片FSDP，把视觉编码移出训练计算图。LongCat称，经过上述优化，该FSDP配置在长序列SFT任务中的单卡吞吐量较DDP基线提升86%。
模型发布后，智东西也第一时间围绕视觉编程和办公任务展开实测。从实测体验上看，模型生成速度较快，在编程和视觉理解上有着不错的表现：前端视觉效果较好，PPT图表丰富，但画面细节、空间与图层关系仍有提升空间，PPT排版也可以进一步优化。
首先，我们让模型制作一段“鹈鹕骑自行车”的HTML动画，LongCat-2.5生成的动画完成度较高，主体突出，画面风格保持一致，细节也比较丰富。不过，鹈鹕的围巾稍微有些歪，画面细节有一定的提升空间。
后面，我们又让LongCat-2.5解读亚马逊第三季度财报，并生成一份带图表的 PPT。
LongCat-2.5搭建的分析框架比较完整，趋势图、结构图和风险提示的搭配也比较自然，模型还根据财报给出投资亮点和风险提示。经初步比对，我们没有发现数据错误。不过，LongCat-2.5生成的PPT部分页面信息密度偏高、部分文本字号偏小，个别结论下得略绝对，排版上还可以再优化调整。
在线体验：
https://longcat.chat
API接入：
https://longcat.chat/platform/product
一、动画自然可交互，任务场景覆盖软件开发与专业办公
LongCat官方团队也给出了几个案例，重点展示了模型在软件开发和专业办公等方面的能力。
在软件开发方面，模型可以生成实时流体模拟与视觉特效生成工具，构建出包含高帧率流体动画的交互式网页。鼠标在界面上点击或拖拽后，可形成墨水流动的动画。
LongCat-2.5也可以开发3D文物展示页面。模型结合WebGL渲染与前端交互技术，生成了三星堆青铜神树的3D模型，用户可以通过旋转、缩放、双击聚焦等方式从不同视角观看文物，展开文物信息面板。
在创意设计和3D场景生成方面，模型能够构建城市沙盘、微缩旋转木马场景和中国传统色彩主题网页，并完成从视觉风格提取、页面布局到交互实现的完整流程。
在办公场景中，LongCat-2.5可以整合多渠道经营数据，梳理数据口径并计算真实利润，进一步根据业务目标分析经营表现和投入产出，生成营销、供货、库存与预算配置建议。
模型还能够分析客户反馈，识别重复问题和后续更新事项，按照问题严重程度、紧迫性与业务目标进行排序，并生成支持点击查看原始反馈的交互式HTML报告。此外，用户还可以上传场地照片、输入自然语言需求，生成包含报名、名单查看和表格导出功能的活动邀请网页。
二、1.6万亿参数，视觉与文本统一为离散Token
LongCat-2.5采用原生多模态架构，将不同分辨率的图像编码为离散Token序列，再与文本Token一同输入主干模型。
在视觉编码环节，LongCat-2.5沿用了LongCat-Next的方法，由语义对齐编码器（SAE）生成原生分辨率特征，再通过残差向量量化RVQ将视觉信息压缩为多层级词表Token，以缓解视觉离散化过程中的信息损失。
模型还引入离线视觉Token预处理机制。固定tokenizer可以提前生成并缓存视觉TokenID，主干模型直接读取预处理后的序列，简化长序列打包与切分，并改善分布式训练时不同任务之间的负载均衡。
LongCat团队称，视觉码本Embedding与主干模型采用联合学习方式。在平均激活参数约30亿的较小规模MoE模型上，视觉与文本Embedding在t-SNE可视化结果中呈交织分布，说明两类信号已经开始在共享语义空间中融合。
三、GitHub PR自动生成可交互环境，长程Agentic RL提升端到端工程
LongCat-2.5围绕长程Agent任务搭建了完整的Agentic RL训练链路，覆盖数据、环境、算法和Harness等方面。
在数据层面，LongCat团队从GitHub公开仓库和内部真实场景中获取代码任务，针对筛选后的PR和关联issue，自动生成可交互环境与可验证测试用例，并通过fail-to-pass和pass-to-pass测试校验任务是否有效。
在RL训练过程中，模型的实时rollout轨迹和任务通过率会被用于二次数据标注与筛选。团队还基于假阴性、假阳性分析，持续修正和扩充F2P测试用例，提升测试对任务功能描述的覆盖率和验证信号的准确性。
针对长程任务中部分轨迹耗时过长的问题，LongCat-2.5采用partial rollout异步训练策略。每个RL step保持一定过采样率，批次凑齐后暂停未完成轨迹，并在下一轮采样时恢复。系统会根据任务通过率动态调整过采样率，同时过滤全对或全错的零梯度样本组，以保证训练过程中的有效梯度。
由于同一条轨迹中的Token可能来自不同历史checkpoint，训练还会面临样本陈旧性问题。为此，LongCat-2.5会裁剪采样阶段与当前训练阶段之间的Logprob差异，并根据历史累计差异调整每个Token的梯度贡献，降低陈旧样本造成的梯度偏置。
在奖励设计上，模型会通过断网、清除git历史和文件缓存等方式限制投机行为，并引入LLM judge监督完整轨迹。对于重复思考、重复调用工具等行为，系统会施加额外惩罚；对于过度思考和思考长度增长过快的问题，则通过重新分配正样本Reward进行抑制。
LongCat团队还采用跨任务、跨Scaffold的混合训练方式，使模型能够适应近似任务和不同Harness。相比冷启动阶段只覆盖少量适配用例，这种训练方式能够进一步提升模型在不同任务环境中的泛化能力。
四、专家组机制覆盖Agent、推理与交互能力
为了将不同领域能力融入主线模型，LongCat-2.5延续了LongCat-2.0的专家组机制，分别优化Agent能力、推理能力和交互体验，覆盖多轮任务执行、单轮推理、细粒度指令遵循以及安全防御等场景。
在MOPD训练中，学生模型会在不同领域任务中在线采样，领域专家则根据学生采样结果的全词表KL散度训练学生模型。
为降低教师轨迹与学生采样之间的分布偏移，LongCat-2.5进一步将教师每轮回复按固定长度切分为per-cut prefix，并将学生模型的自采样长度限制在相应范围内。训练时，同一教师轨迹的多个Prefix会被分组送入同一推理引擎，共享Prefill计算和KV Cache。
LongCat团队还以教师轨迹为树干、学生采样结果为分支构建训练树，结合Tree Attention，通过一次前向和反向计算完成同组样本的OPD训练，从而减少教师轨迹带来的重复计算与存储开销。
五、围绕超节点架构重新设计并行策略，单卡吞吐量较基线提升86%
LongCat-2.5在预训练和监督微调阶段均采用长序列训练。针对国产AI加速器单卡显存不足80GB的限制，训练系统围绕超节点架构重新设计并行策略。
LongCat-2.5根据超节点拓扑配置并行组，FSDP和EP并行度均与单个超节点的卡数保持一致，CP并行度不超过该卡数，使参数AllGather、专家TokenAlltoAll和KVAllGather尽可能在超节点内部的高带宽互联域中完成。
系统还通过混合词表并行，对基础词嵌入层、LMHead和N-gram嵌入层进行分片，进一步降低显存占用。针对大规模专家并行中的负载不均衡，LongCat-2.5基于MoonEP引入阈值触发的动态专家复制机制，仅在负载不均衡达到预设阈值时启用复制，以平衡通信成本与负载收益。
通过提前生成图像Token，训练计算图不再包含视觉编码器。系统可以根据已知的图像Token数量，提前估计数据打包和打散时的多模态负载比例，使不同数据并行Rank上的模态占比更加接近。
LongCat-2.5还引入双重分片FSDP。模型参数在超节点内的FSDP域中分片，FP32主权重和优化器状态则通过HSDP副本进一步跨FSDP域分片，在保持参数通信效率的同时减少常驻显存。系统通过批量执行梯度操作、合并指标归约、CPU 绑核和 NUMA 感知调度，降低主机侧开销。
LongCat称，经过上述优化，该FSDP配置在长序列SFT任务中的单卡吞吐量较DDP基线提升86%。
结语：原生多模态模型开始向长程工程执行推进
从LongCat-2.5的技术路线看，模型升级重点已经从单纯扩大参数规模，转向视觉统一建模、长程Agent训练、训推一致性和国产算力适配等系统工程。
尤其是在复杂代码库开发、3D页面生成和多轮工具调用等任务中，模型能否持续执行、验证结果并根据失败轨迹修正策略，将成为衡量Agent能力的重要指标。LongCat-2.5把原生视觉理解与长程Agent训练结合起来，也意味着多模态模型正在从“看懂内容”进一步走向“完成任务”。
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。
Intellectual (WeChat Official Account:zhidxcom)
YANG Kyung-Li
Editor, Lee Shui-qing
Shigeki reported on November 2 that today, the LongCat team publish of the United States was looking at the original multi-modular model LongCat-2.5 for the AI agent mission.
The model, with a total of parameter count of 1.6 trillion, activates approximately 48 billion parameters per Token, focusing on strengthening native visual understanding and AI agent programming capabilities based on LongCat-2.0.
LongCat-2.5 leads models such as Claude-Opus-5, GPT-5.6-Sol in several benchmark evaluations such as AutomationBech, the Generic Agent mission benchmark, and LongClawBech, published by the United States Corps; LongCat-2.5 is divided between 38.5 and over GPT-5.6-Sol, but still lags behind Claude-Opus-5 in the terminal operating benchmark Term Bench v.4.
The LongCat-2.5 long-sequence training introduced more long-range track data and enhanced end-to-end engineering capabilities of models in complex code library development, complete project construction, and so forth, through large-scale long-range intensive learning. At the same time, LongCat-2.5 transforms visual signals into multiple layers of discrete Token, aligning visual and text input with the same backbone model, further supporting visual programming, creative design and 3D content generation.
In the face of the VRAM limit of a single card with a national accelerator of less than 80GB, LongCat-2.5 was reordered to the FSDP, EP, CP parallels around the hypernodes, superseding the syntax, MoonEP dynamic expert copy and double FSDP, moving the visual code out of the training calculations. LongCat stated that, following the above-mentioned optimization, the FSDP configuration in the long series SFT missions increased single-calorie throughput by 86 per cent over the DDP baseline.
Following the model publish, intellectuals were first measured around visual programming and office assignments. From an empirical experience, models are produced faster and perform well in programming and visual understanding: the front end is better visualized, the PPT charts are rich, but there is room for improvement in graphic details, spatial and layer relationships, and the PPT layout can be further optimized.
First of all, we have modeled animated animated HTML by a "piper-cycling" and LongCat-2.5 produced a high degree of animation, a high profile, consistent image style and more detailed. However, the pelican scarf is slightly crooked, and there is some room for improvement in the details.
And then we got LongCat-2.5 to read the Amazon 3rd Quarter and produce a PPT with charts.
LongCat-2.5 has a more complete analytical framework and a more natural mix of trend maps, charts and risk tips. The model also provides investment highlights and risk tips based on financial statements. After a preliminary comparison, we found no data errors. However, the PPT portion of LongCat-2.5 is produced with a high level of information density and a small number of text fonts, with a slightly absolute number of individual conclusions, which can be fine-tuned.
Online experience:
https://longcat.chat
API access:
https://longcat.chat/platform/product
I. Animation is naturally interactive, mission scenario covers software development and professional office
The official LongCat team also gave several examples highlighting the capabilities of models in software development and professional office.
In terms of software development, models can generate real-time fluid simulations and visual effects generation tools to build interactive web pages containing high-spectrum fluid animation. Click or drag the mouse on the interface to form an animation of ink flow.
LongCat-2.5 can also develop a 3D exhibition page. The model, combined with WebGL rendering and front-end interaction techniques, produces a 3D model of a three-star pile of brass trees, which allows users to view cultural objects from different perspectives through rotation, scaling, double-click focus, etc., and expands the cultural objects information panel.
In terms of creative design and 3D scenario generation, models are able to construct city sandpads, micro-scaling wood horse scenes and traditional Chinese colour theme pages and complete the full process from visual style extraction, page layout to interactive realization.
In the office scene, LongCat-2.5 consolidates multi-channel operating data, collates data calibres and calculates real profits, further analyses business performance and input output against business objectives, and generates marketing, supply, inventory and budget allocation proposals.
The model also allows for the analysis of client feedback, the identification of duplication issues and subsequent updates, the sequencing of issues according to the severity, urgency and operational objectives, and the generation of interactive HTML reports that support clicking on original feedback. In addition, users can upload photographs of the site, enter natural language requirements, and generate pages of invitations to events that include registration, list viewing and table export functions.
II. 1.6 trillion parameters, visual and text alignment to discrete Token
LongCat-2.5 uses a native multimodular structure, encodes different resolution images into discrete Token sequences and then enters the backbone model with the text Token.
In the visual coding chain, LongCat-2.5 follows the LongCat-Next method, where the semantic alignment encoder (SAE) produces the original resolution feature and compresses visual information to Token, a multi-level wordsheet, through the disability vector quantizationRVQ, to mitigate the loss of information during visual fragmentation.
The model also introduces the offline visual Token pre-treatment mechanism. A fixed tokenizer can generate and cache visual TokenID in advance. The backbone model directly reads preprocessed sequences, simplifys long sequence packs and cut points and improves the load balance between different tasks during distributed training.
According to the LongCat team, the visual code book Embeding and the backbone model use a joint learning approach. On the smaller MoE model, with an average activation parameter of about 3 billion, visual and text Embeding presented a woven distribution in the t-SNE visualization results, indicating that two types of signals had begun to merge in shared semantic spaces.
III. GitHub PR auto-generated interactive environment, long-range Agenic RL upend to end project
LongCat-2.5 complete Agentic RL training links around the long-range Agent mission covering data, environment, algorithms and Harness.
At the data level, the LongCat team obtains code assignments from the GitHub open warehouse and from the internal reality scene, automatically generate interactive environments and verifiable test examples for the filtered PR and associated issue, and test the validity of the check assignments through fail-to-pass and pass-to-pass.
During RL training, real-time rollout tracks and mission pass rates of models are used for secondary data labelling and filtering. The team also continuously revises and expands the F2P test to enhance the coverage and accuracy of the mission function description of the test, based on false negatives and positive analysis.
LongCat-2.5 uses a partial rollout training strategy to address the excessive length of the trajectories in part of a long-range mission. Each RL step maintains a certain sampling rate, suspends the incomplete trajectory after the batch has been completed and recovers at the next round of sampling. The system adjusts the sampling rate to the mission's rate dynamics, while filtering the zero gradient group sample, which is correct or all wrong, to ensure an effective gradient during training.
Since Token in the same trajectories may come from different historical checkpoints, the training also faces the problem of sample obsolescence. To this end, LongCat-2.5 will tailor the Logprob difference between the sampling phase and the current training phase and adjust the gradient contribution per Token based on cumulative historical differences to reduce the gradient bias caused by old sample.
In incentive design, models limit speculation by cutting off nets, removing git history and document cache, and introduce LLM jurisprudence to monitor the full trajectory. The system imposes additional penalties for re-thinking, duplicate call tools, etc.; the over-thinking and over-thinking of excessive length growth is suppressed by redistribution of positive sampleReward.
The LongCat team also uses a cross-mission, cross-Scaffold hybrid training to adapt the model to similar tasks and different Harness. Compared to the cold start-up phase, which covers only a small number of suitable applications, this training will further enhance the model ' s ability to extend across missions.
IV. REPRESENTATION OF THE EXPERT GROUP MECHANISM TO Agent, ELIMINATION AND INTERACTIVE CAPACITY
In order to integrate capacities in different fields into the main line model, LongCat-2.5 has continued the LongCat-2.0 expert group mechanism, which optimizes the capacity of Agent, the ability to reason and the experience of interaction, respectively, to cover multiple rounds of mission execution, single-wheel reasoning, follow-up to fine particle size directives and security defences.
In MOPD training, student models are sampled online in different fields of assignment, while field specialists train student models based on the full word list of student sampling results, KL diffusion.
LongCat-2.5 further divides teachers ' responses per round by fixed length into per-cut prefix, and limits the length of self-sampling of student models to the same extent. During the training, several Prefix tracks of the same teacher were sent to the same reasoning engine, sharing Prefille calculations and KV Cache.
The LongCat team also constructed training trees for branches using the teacher tracks as a trunk and student sampling results, and completed the OPD training for the same group of sample through a forward and reverse calculation in conjunction with the Tree Attention, thereby reducing the double counting and storage costs associated with the teacher tracks.
V. Re-engineered parallel strategies around the hypernomous structure, with single-calorage vomiting up 86 per cent from baseline
LongCat-2.5 uses long-series training at the pre-training and supervisory fine-tuning stages. In response to the limitation of the National Production Accelerator Single Card VRAM of 80GB, the training system re-engineered parallel strategies around the hypernomic architecture.
LongCat-2.5 is configured in parallel groups based on hypernodes, FSDP and EP parallels are consistent with the number of cards for a single hypernodes, and CPs do not exceed that number, so that the parameters AllGather, experts TokenAlltoAll and KVAAllGather are completed in the high bandwidth interconnection within the hypernodes to the extent possible.
The system also uses a hybrid vocabulary to further reduce the VRAM occupancy by splitting the basic words embedded layer, the LMhead and the N-gram embedded layer. LongCat-2.5 is a dynamic expert reproduction mechanism triggered by the introduction of a MoonEP threshold for the simultaneous load imbalance of large-scale experts, only to enable replication when the load imbalance reaches the predefined threshold to balance communication costs against load gain.
By producing the image Token in advance, the training calculator no longer contains visual encoders. Based on the number of images known as Token, the system can estimate in advance the multi-modular load ratio at the time of data packing and disassembly, bringing the model share of different data in parallel Ranks closer.
LongCat-2.5 also introduced a dual FSDP. Model parameters are split in the FSDP domain within the hypernodes, while the FP32 sovereign weight and optimizer state further crosses the FSDP domain by means of a HSDP copy, reducing permanent presence in VRAM while maintaining the efficiency of the parameter communication. The system reduces the cost of the side of the host by batches of gradient operations, consolidation of indicator alignment, CPU binding and NUMA sensoring.
LongCat stated that, following the above-mentioned optimization, the FSDP configuration in the long series SFT missions increased single-calorie throughput by 86 per cent over the DDP baseline.
Conclusion: Prototype multi-modular models begin to advance towards long-range engineering
From the technical route of LongCat-2.5, the focus of model upgrades has shifted from simple expansion of parameters to systems engineering such as visual integrated modelling, long-range Agent training, training consistency and national production compatibility.
Especially in complex code library development, 3D page generation and multiple rounds of tool call missions, the sustainability of the model, validation of results and correction of strategies based on failed trajectory will be important indicators of Agent capacity. LongCat-2.5, which combines native visual understanding with long-range Agent training, also means that multi-modular models are moving further from “understanding content” to “manufacturing”.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.