产品Product / 混合算力网关Hybrid-compute gateway
本地 GPU 与云端模型，
走同一道门Local GPUs and cloud models, through one door
模型网关部署在企业自己的环境里。本地模型和云端模型接到同一个入口，出现在员工同一个模型列表里。每一次请求都按人校验、限额和计量。
The model gateway runs in your own environment. Local and cloud models connect to one entry point and appear in one model picker for every employee. Every request is checked, limited and metered per person.
一次请求的路径The path of one request
员工在客户端里发出一次请求，网关按下面的顺序处理。
When an employee sends a request from the client, the gateway handles it in this order.
- 客户端发起Client sends
请求带着这位员工自己的网关密钥。
The request carries the employee's own gateway key.
- 校验密钥Check the key
员工被停用时，密钥已被封，请求返回 401。
If the employee is disabled, the key is blocked and the request gets 401.
- 检查额度Check the quota
预算、每分钟请求数、每分钟 token 数和可用模型。超出速率返回 429。
Budget, requests and tokens per minute, and allowed models. Over the rate limit, it returns 429.
- 路由到模型Route
本地模型在企业的 GPU 上推理，云端模型转发到你配置的供应商。
Local models run on your GPUs; cloud models go to the provider you configured.
- 按人计量Meter
记录员工、模型、token、花费、状态和耗时，不记录对话内容。
Records employee, model, tokens, spend, status and latency. No message content.
- 返回结果Respond
结果回到员工的客户端。
The answer returns to the employee's client.
本地与云端，各取所长Local and cloud, each where it fits
统一网关One gateway| 本地模型Local models | 云端模型Cloud models | |
|---|---|---|
| 在哪里推理Where inference runs | 企业自己的 GPU 主机。代码和对话不出企业网络。Your own GPU host. Code and conversations stay inside your network. | 你配置的云端供应商。The cloud provider you configure. |
| 怎么接入How it connects | 在控制台的本地算力页启动和停止。Start and stop it on the Console's Local compute page. | 在模型与供应商页填写 https 接口地址、协议和密钥。Add an https base URL, protocol and key on the Models and providers page. |
| 计量与额度Metering and quota | 两者一样：按人计量，按人限额，同一套审计。可以给每个模型设置计价，用于计算额度。The same for both: metered per person, limited per person, one audit trail. Each model can carry its own price for quota accounting. | |
员工只看到一个模型列表Employees see one model picker
员工不用关心模型在哪里运行。客户端里的模型列表来自控制台的模型目录，员工能调用哪些模型由他的可用模型决定。管理员改目录，不需要重新打包客户端。
Employees never need to know where a model runs. The picker in the client comes from the Console's model catalog, and each employee's allowed models decide what they can call. Change the catalog without shipping a new client.
- 本地模型和云端模型并列在同一个列表里，用员工看得懂的显示名。
- 目录里有、员工又被允许的模型，员工才能正常使用。
- Local and cloud models sit side by side, under display names employees understand.
- An employee can use a model when it is in the catalog and on their allowed list.
按人管控，立即生效Per-person control that takes effect at once
-
每人一把密钥A key for each person
员工第一次登录时，控制台在网关里为他开账号并发一把只属于他的密钥。
At an employee's first sign-in, the Console opens a gateway account and issues a key that belongs to them alone.
-
额度跟人走Quotas follow the person
额度挂在员工身上。删除密钥、重新登录都不会重置额度。
Quotas belong to the employee. Deleting a key or signing in again never resets them.
-
停用即封禁Disable means blocked
停用员工后，网关封禁他的密钥，下一次请求就失败。重新启用后恢复。
Disable an employee and the gateway blocks their key, so the very next request fails. Enable them to restore access.
-
定期对账Regular reconciliation
控制台每 300 秒和网关对账一次。网关暂时不可用时，停用操作在网关恢复后补上。
The Console reconciles with the gateway every 300 seconds. If the gateway is briefly down, a pending disable is applied when it comes back.
把你的算力接进来Bring your own compute
看看本地 + 云和纯云两种部署方式，哪种适合你。
See whether local plus cloud or cloud only fits your team.