从公网入口，到模型执行。
From public ingressto model execution.
从公网入口，到模型执行。
From public ingressto model execution.
EggyByte Technology 纵向构建公网 Edge、分布式控制与状态系统、YaoBox AI Runtime、Data / Model Gateway 与模型工程，让网络、状态、执行、数据和智能保持在同一套系统边界内。
EggyByte Technology vertically engineers public edge, distributed control and state, the YaoBox AI runtime, Data / Model Gateway and model engineering—keeping network, state, execution, data and intelligence inside one systems boundary.
公网入口保持轻，业务权威留在核心。
Keep the edge small.Core holds authority.
公网入口保持轻，权威留在核心。
Keep the edge small.Core holds authority.
规划中的公网入口与 OVH、ISIF 等基础设施服务方协同部署具备 10Gbps 网络能力的 Edge 节点。Controller 主动连接 Edge，并以 QUIC v1、TLS 1.3 与固定 Ed25519 身份认证；Edge 仅处理公网 TLS、连接接入与有序字节流转发。
The public ingress plan uses edge nodes with 10Gbps networking capability together with infrastructure providers including OVH and ISIF. Controllers initiate the connection to each edge over QUIC v1 and TLS 1.3 with fixed Ed25519 identities; the edge remains limited to public TLS, connection ingress and ordered-byte-stream forwarding.
按角色横向扩展，不重写系统拓扑。
Scale each role.Keep the topology.
按角色横向扩展，不重写系统拓扑。
Scale each role.Keep the topology.
Yaoledge 将无状态 Controller、Model Gateway、Data Gateway、Linux NodeSession 与 Storage Box 拆成独立扩展域。任意数量 Controller 共享 FoundationDB / Yaostate 权威 Product State；后台工作按固定 shard 与 lease 并行推进，不依赖单一全局 Leader。
Yaoledge separates stateless controllers, Model Gateway, Data Gateway, Linux NodeSessions and Storage Boxes into independent scale domains. Any number of controllers share authoritative Product State in FoundationDB / Yaostate, while background work advances through fixed shards and leases instead of a single global leader.
为 AI 执行，重做 Runtime。
A runtime rebuiltfor AI execution.
为 AI 执行，重做 Runtime。
A runtime rebuiltfor AI execution.
Node 主动向 Controller 拉取命令，不暴露公网或私网入站 RPC。YaoBox 直接组合 Linux cgroup v2、user / network namespace、EROFS、OverlayFS、XFS project quota 与 crun，为每个任务建立独立 PID、网络、用户、cgroup 与可写文件系统边界。
Nodes pull commands from controllers and expose no public or private inbound RPC. YaoBox directly composes Linux cgroup v2, user / network namespaces, EROFS, OverlayFS, XFS project quotas and crun, giving every task explicit PID, network, user, cgroup and writable-filesystem boundaries.
计算与智能，各自扩展。
Scale compute andintelligence independently.
计算与智能，各自扩展。
Compute and intelligencescale independently.
Controller 负责业务授权、状态与计费语义；无状态 Model Gateway 负责请求 staging、模型路由、Provider Credential、上游执行、流式解析与标准化。Node 只提交绑定当前 Task 的标准请求，不接触真实 Provider Endpoint 或 Credential；同一层可继续接入 Ember、Nova 与其他自研模型。
Controllers own product authorization, state and billing semantics; the stateless Model Gateway owns staging, model routing, provider credentials, upstream execution, stream parsing and normalization. Nodes submit standardized requests bound to the current Task and never receive provider endpoints or credentials; the same plane can progressively host Ember, Nova and other in-house models.
不掩盖失败。恢复必须确定。
Never hide failure.Deterministicrecovery.
不掩盖失败。恢复必须确定。
Never hide failure.Recoverdeterministically.
模型调用采用 at-most-once Provider effect：Controller 先把请求持久化为 POSSIBLE 并绑定唯一 GatewayInstance，Gateway 才能发送第一个 Provider 请求字节。一旦请求可能到达上游，其他 Gateway 不自动重放；不确定结果通过 readback 与 reconcile 收敛。
Model calls use an at-most-once provider-effect protocol: the controller first persists the request as POSSIBLE and binds one GatewayInstance before the gateway may send the first provider byte. Once the request may have reached upstream, another gateway cannot replay it automatically; uncertain outcomes converge through readback and reconciliation.
模型本身，也是系统工程。
The model itselfis a systems problem.
模型本身，也是系统工程。
The model itselfis a systems problem.
模型工程把架构、数值精度、GPU 通信、长上下文执行与 Serving 作为同一系统联合设计：现有路线覆盖 NVFP4-first 数值训练、MXFP8 敏感 attention path、Granular MoE、Gated DeltaNet、hierarchical sparse attention 与 StateFlow 长上下文 pipeline。
Model engineering co-designs architecture, numerical precision, GPU communication, long-context execution and serving as one system. The current route spans NVFP4-first training, MXFP8-sensitive attention paths, Granular MoE, Gated DeltaNet, hierarchical sparse attention and the StateFlow long-context pipeline.
一套系统，连接产品与真实商业链路。
One foundation.Real products.Real distribution.
一套系统，连接产品与真实商业链路。
One foundation.Real products.Real distribution.
Yaoledge 面向分布式 AI 执行；Yaoedge 与后续产品复用同一网络、状态、执行和智能底座。基础设施与 OVH、ISIF 协作；支付与商业基础设施接入支付宝；流量、广告与用户触达侧连接优量汇、穿山甲与快手。
Yaoledge targets distributed AI execution; Yaoedge and subsequent products reuse the same network, state, execution and intelligence foundation. Infrastructure work connects with OVH and ISIF; payment and commerce infrastructure integrates Alipay (支付宝); traffic, advertising and user reach connect with 优量汇, Pangle (穿山甲) and Kuaishou (快手).
