可以不出。要私有化的话，系统装在你们自己的机房，有三种形态：一台单机或一体机、机房 GPU 集群、完全断网的内网离线版。离线版用安装包交付，模型、知识库和日志都留在内网，断网也能用。

我们按你们的业务场景来做，用的是你们自己的资料和流程，做完按结果验收。应用、知识库、算力和运维都是同一个团队负责，出了问题找我们就行。可以先用云端试，效果对了再把同一套东西搬到你们机房，业务不用重做。接口统一，方便接你们现有的系统。

先挑一个最值得做的场景，7–15 天做出能演示、能试用的版本，可以先在云上跑。交付范围和验收标准在开始前就定好。看过效果，再决定放云上还是本地、要不要推到别的场景。

这是合作的第一步。我们一起看场景、数据、合规要求和现有系统，已有 GPU 的也一起看，然后给出先做哪个场景、放云上还是本地，判断是否必须用本地算力。之后再进入 7–15 天的试点。

云端按用量或按席位收费。本地部署有三档硬件预算参考：入门单机约 100 万（一台 RTX 5090 八卡机），可商用分池约 200–300 万（多台 5090 八卡集群），数据中心旗舰约 400–600 万（H100 / H200 级节点）。这些是行情参考，不锁价，硬件以下单时的询价为准。这三档只是硬件费用，安装、调试、压测和软件不含在内，另外报价。合同里会写明同时会话数和月度额度。

最低一档是一台 RTX 5090 八卡机（大内存），硬件预算约 100 万档，行情参考；安装、调试、压测和软件另外报价。这一档以文本为主：对话、摘要、写作和知识库都能跑，图文能力减半，视频只到演示级。软件用 vLLM、LiteLLM、Open WebUI 和向量库。要更高并发、做图文和视频，就上多机集群或 H100 / H200 节点。

已有 GPU、想把机器用起来的团队，也在我们的服务范围里。在 AI 企业方案落地咨询阶段，先看你们现有的机器和系统，再定方案。部署后多张卡统一调度，对内只开一个接口，主流开源和商用模型都能接。

看效果、成本和合规要求。想先快点看到效果，用云端，开通就能用。小团队、门店、科室想做私有化试点，用一台单机或一体机；并发高、要稳定推理，用机房 GPU 集群；政企、金融、制造、医疗这类数据敏感的行业，或者不能联网的等保、涉密环境，做内网离线。也可以先在云上跑，效果确认后整体搬到本地，业务不用重做。

AI 负责识别、查资料和起草：比如识别询盘里的客户、产品和意图，对上产品资料，写好回复草稿，并注明引用了哪份资料。报价这类高风险操作，要人审批后才发出。知识库有版本记录，每条回答都标出处，事后查得到。

还是我们。账号权限、用量统计、安全审计和培训都在交付范围里，上线后接着改。本地部署的话，GPU 占用、调用量、失败率和响应时间都能看到。

个人项目我们不接，只做企业客户。如果只想要一个演示用的聊天机器人，也不用找我们。我们的做法是先把一个真实业务场景跑通，再按结果验收。

No. For private deployment the system is installed in your own data center, in one of three forms: a single server or appliance, a GPU cluster, or a fully air-gapped offline build. The offline build ships as an installer; models, knowledge base and logs stay on your network and everything works without an internet connection.

We build around your use case, with your own documents and processes, and you sign off on results. Apps, knowledge base, compute and operations are all handled by one team, so there's one place to go when something goes wrong. You can start in the cloud and, once it works, move the exact same system into your data center without rebuilding. A unified API makes it easy to connect to the systems you already run.

We pick the use case most worth doing and deliver a version you can demo and try within 7–15 days, which can run in the cloud to start. Scope and acceptance criteria are agreed before work begins. After you've seen results, you decide on cloud vs on-prem and whether to expand.

It's the first step. Together we look at your use cases, data, compliance requirements and existing systems, including any GPUs you already own. We then recommend which use case to start with, whether it belongs in the cloud or on-prem, and whether on-prem compute is really needed. A 7–15 day pilot follows.

Cloud is billed by usage or by seat. For on-prem there are three hardware budget references: an entry server at about RMB 1M (one 8× RTX 5090 machine), production pools at about RMB 2–3M (a cluster of 8× 5090 servers), and a data-center flagship at about RMB 4–6M (H100 / H200-class nodes). These are market references, not fixed prices; hardware is quoted at order time. All three figures are hardware only. Installation, tuning, load testing and software are not included and are quoted separately. The contract states concurrent sessions and monthly quota.

The entry tier is one 8× RTX 5090 server with large memory, with a hardware budget of around RMB 1M (market reference; installation, tuning, load testing and software are quoted separately). It's mainly for text: chat, summaries, writing and knowledge base all run well, image capacity is halved, and video is demo-level only. The software stack is vLLM, LiteLLM, Open WebUI and a vector database. For higher concurrency, images and video, you move up to a multi-node cluster or H100 / H200 nodes.

Yes. Teams that already own GPUs and want to put them to work are part of what we do. During the assessment we review your current machines and systems before settling on a plan. After deployment, all GPUs are scheduled as one pool behind a single internal API, and mainstream open-source and commercial models can be plugged in.

It comes down to results, cost and compliance. To see results quickly, start in the cloud. For a small team, branch or department piloting private deployment, use a single server or appliance. For high concurrency and production inference, use a GPU cluster in your data center. Data-sensitive sectors such as government, finance, manufacturing and healthcare, or regulated environments with no internet access, go air-gapped. You can also start in the cloud and move the whole system on-prem once it's proven, without rebuilding.

AI identifies, looks things up and drafts: for example, it reads an inquiry, works out the customer, product and intent, matches the right product docs and writes a draft reply noting which documents it used. High-risk actions like sending a quote need human approval first. The knowledge base is versioned and every answer cites its source, so everything can be traced afterwards.

We do. Access control, usage reporting, security audit logs and training are part of the delivery, and we keep improving it after launch. On-prem, you can see GPU utilization, call volume, failure rate and response time.

We don't take personal projects; we work with businesses only. If all you need is a chatbot for a demo, you don't need us either. Our approach is to get one real business workflow working first and be judged on the results.