背景
美团龙猫发布了新一代原生多模态模型 LongCat-2.5，面向长程智能体任务。模型总参数量达 1.6 万亿，每个 token 激活约 480 亿参数，属于典型的大规模混合专家结构。这是在 LongCat-2.0 的基础上迭代而来，重点方向是强化原生视觉理解和智能体编程能力。
长程智能体任务指的是需要模型在较长时间跨度、较多步骤中保持连贯决策的工作，比如复杂代码库的开发和完整项目的构建。这类任务对模型的要求和单轮问答完全不同,需要模型能记住之前的操作、判断中间结果是否正确，并根据反馈调整后续步骤。LongCat-2.5 的训练方式就是围绕这个目标设计的。
目前模型已经可以在 longcat.chat 在线体验，API 接入入口也已开放，GitHub 和 HuggingFace 上的开源版本官方表示即将上线。
核心原理
LongCat-2.5 在训练阶段加入了更多长程轨迹数据，并配合大规模长程强化学习,目的是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。官方还提到，围绕数据合成、环境构建和跨 Harness 混合训练，这条训练链路在持续扩展。
一个值得关注的做法是，龙猫团队把真实 Agent 运行时产生的失败轨迹转化成可复现、可验证的训练任务，再针对模型暴露出的能力短板补充训练数据。这种从失败中反向构造训练样本的方式，比单纯依赖人工标注数据更贴近真实使用场景中会遇到的问题。
在多模态处理上，LongCat-2.5 把视觉信号表示为多层级离散 token，让主干模型用统一的 token 形式同时处理视觉和文本输入。这种设计使得原生视觉理解和智能体编程能力可以结合起来，模型在 Visual Coding（结合视觉信息进行代码生成或调整）上表现突出，并支持创意设计、3D 相关任务。
支撑这些能力的是持续演进的基础设施：官方提到在国产算力上优化了长上下文训练和 W4A4 推理，还通过训练与推理之间的数值一致性优化,以及长上下文 rollout 优化，来保证大规模长程强化学习能够高效运行。
局限
素材中没有给出 LongCat-2.5 在具体基准测试上的分数，也没有披露训练数据规模、训练时长或硬件集群规模等细节，这些信息目前无法确认。
模型的开源版本尚未上线，GitHub 和 HuggingFace 页面官方标注为“即将上线”，具体开放时间和开源协议都还不明确。API 接入页面目前能查到入口，但计费方式、调用限制、支持的上下文长度等参数素材中没有提及。
“W4A4 推理”说明模型支持 4-bit 权重和 4-bit 激活的量化推理，这能降低显存占用和推理成本，但量化之后的精度损失、实际推理速度提升幅度，素材中也没有给出数字，这里不做推测。
对开发者的意义
对于需要处理长流程智能体任务的开发者，LongCat-2.5 提供了一个新的候选模型，尤其是涉及代码库级别的开发、多步骤项目构建的场景。原生多模态的设计意味着视觉输入和文本输入可以在同一套 token 体系下处理，如果产品里有结合图片或设计稿生成代码的需求，这个方向值得关注。
在正式接入前，建议先通过在线体验入口测试模型在自己业务场景下的实际表现，再决定是否通过 API 接入。开源版本上线后,本地部署和微调的可行性会更清晰，目前只能等待官方后续动作。
量化推理支持对于成本敏感的部署场景是个积极信号，但具体收益需要等实测数据或官方公布的性能报告。
原文地址：tools.wflynn.cn/aitools/tre…
关注公众号「YourTools」，每天推送 GitHub 趋势与开源工具速报：
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。
Background
LongCat-2.5, a new generation of original multi-modular models for the Long-Range AI agent mission. The model totals parameter count to 1.6 trillion, each token activates about 48 billion parameters, which are typical of large-scale hybrid specialist structures. This is based on LongCat-2.0, with emphasis on strengthening native visual understanding and AI agent programming capabilities.
The long-range AI agent mission refers to the need for a model to maintain coherent decision-making over a longer time horizon and in more steps, such as the development of complex code libraries and the construction of complete projects. The requirements of this type of task are completely different from those of a single question and answer cycle, requiring models to be able to remember previous operations, to judge whether intermediate results are correct and to adjust subsequent steps based on feedback. LongCat-2.5 training methods are designed around this goal.
The model is now available online at Longcat.chat, the API access is open, and the open-source version on GitHub and HuggingFace officially indicates that it is on line.
Core principles
LongCat-2.5 Additional long-range track data were added to the training phase and complemented by large-scale long-range intensive learning aimed at upgrading the end-to-end engineering capabilities of the model in complex code library development, complete project construction, etc. Officially, it was also mentioned that the training links around data synthesis, environmental construction and cross-Harness training are expanding.
It is a matter of concern that the dragon-cat team converts the failure trajectory generated by the real Agent running into a reversible, verifiable training mission, supplemented by training data on the capability panels revealed by the model. This way of training sample in reverse construction from a failure is closer to the real use scene than simply relying on manual tagging data.
In multi-modular processing, LongCat-2.5 expresses the visual signal as multi-level separation token, allowing the backbone model to process visual and text input simultaneously in the same token form. This design allows for a combination of native visual understanding and AI agent programming capabilities, with models that feature prominently on Vision Coding (coding or adjusting with visual information) and support creative design, 3D-related tasks.
These capacities are underpinned by a continuous evolution of infrastructure: official references to the optimization of long-term context training and W4A4 reasoning in the country ' s productive capacity, as well as by the optimization of numerical consistency between training and reasoning, and long-term context rollout to ensure that large-scale intensive learning can function efficiently.
Limits
asset did not give LongCat-2.5 scores for specific benchmark tests, nor did it disclose details of the size of training data, the length of training or the size of hardware clusters, which could not be confirmed at this time.
The open-source version of the model is not yet online, and the GitHub and HuggingFace pages are officially marked as “on-line”, and the specific opening hours and open-source protocols are not clear. The API access page currently allows access to the portal, but parameters asset, such as the method of billing, call limits, and context length of support, are not mentioned.
The W4A4 reasoning indicates that the model supports the 4-bit weight and 4-bit activated quantization reasoning, which can reduce the cost of VRAM occupancy and reasoning, but the precision loss, the increase in the actual reasoning speed after quantization, and the figures given in asset are not speculative.
Meaning to developers
LongCat-2.5 provides a new candidate model for developers who need to deal with long pipelineZ42Z tasks, particularly in relation to the development of the code library level, multistep project construction. The design of original multimodulars means that visual input and text input can be handled under the same token system, and this direction is of interest if there is a need in the product to combine a picture or a draft generation code.
Before formal access, it is recommended that the actual performance of the online experience portal model in its own business landscape be tested before deciding whether to access through the API. Once the open-source version is online, the feasibility of local deployment and fine-tuning will be clearer and will now await official follow-up.
The quantization reasoning supports a positive signal of a cost-sensitive deployment scenario, but the specific benefits require such measurements or officially published performance reports.
Original address:tools.wflynn.cn/aitools/tre...
Following the WeChat Official Account "YourTools", GitHub's trend and open source tool is sent daily:
Auto-fetched and translated by AI Atlas. All rights belong to the original author.