XB EDGE · 2026 · PRIVATE EDITION
EDGE COMPUTE · PRIVATE LLM
XB EDGE
XB EDGEEdge Compute · Private LLM
Fullscreen 中文 / EN Book a Demo +86 158 8199 1999
EDGE
Edge Compute · Private LLM

Run the model
inside your own data center

A regional edge compute network plus on-premise LLM deployment: data never leaves the plant, works with the internet unplugged, ready out of the box. We hand the AI capability to the enterprise instead of renting you an API.

XB EDGE assistant
14BLOCAL MODEL
190TOKENS / SEC
0DATA EGRESS
11 nodes live · 300+ county service points (planned)
11 edge nodes 190 tok/s local inference · off-grid 14B on-premise model 0 data egress · demo with the cable unplugged SCROLL ↓
11nodesRegional Edge Nodes
190tok/sLocal Inference
14BOn-Premise Model
0egressData Egress
01 — PLATFORM

Edge compute network
Put the compute next to the enterprise

Instead of competing with hyperscale data centers in big cities, we place edge nodes along the industrial chain: low latency, low cost, direct green power, serviceable within 10 km.

Regional nodes · inference nearby

Nodes are deployed inside parks and plants. Inference latency stays at single-digit milliseconds, with no dependency on public bandwidth and no outage when the internet goes down.

LOW LATENCYON-SITE

Heterogeneous scheduling

CPU and GPU are scheduled together, and model size is matched to the task: 3B for classification, 7B for conversation, 14B for long-form output.

SCHEDULERGPU POOL

Green power · 30% lower cost

Built alongside local renewable generation. Compute and power are coordinated so the energy cost per token undercuts public cloud.

GREEN POWERCOST
02 — PRIVATE LLM

Private LLM
Data stays in, capability grows in-house

Model, corpus, knowledge base and applications are all deployed on the enterprise intranet. We do not touch customer data — we deliver capability.

01

Private knowledge base

Process documents, incident cases and regulatory standards become a dedicated enterprise RAG, with citations and traceability in every answer.

RAG / CITATION
02

Vertical fine-tuning

Incremental training on industry corpora so the model understands your jargon, process steps and risk points.

FINE-TUNE
03

Works offline

The whole stack is local. No external API dependency — when the internet is cut, the AI keeps working.

OFFLINE
04

Domestic OS support

Delivered for Windows and for Kylin / UOS, meeting government procurement and classified-network requirements.

Kylin / UOS
05

Managed operations

Model updates, knowledge increments and fault alerts are hosted remotely. The enterprise needs no in-house AI team.

MANAGED
Organized by business flow

Not a Q&A box — a process card

Every knowledge item states its position in the business flow, the role and documents involved, the workflow steps, the key control points and common deviations — embedded directly into procurement, production, QC, warehousing, sales, service and finance.

Dual-model routing

Local + cloud, switching automatically

Classified, offline and real-time queries run on the local model (3B/7B/14B, 1–3 s, no data leaves the plant); complex reasoning and long documents go to the cloud; when the cloud is unreachable the system degrades to local automatically and the business never stops. Both share one knowledge base, so the answers stay consistent.

03 — INDUSTRY

Six industries
Vertical scenarios already in production

Not a thin wrapper over a general model — solutions shipped with industry corpora and compliance requirements built in.

Chemical safety

Permit-to-work, hazardous chemical storage and leak-emergency Q&A, citing mandatory standards such as GB 30871.

GB 30871SOP

Port & logistics

Dangerous-goods handling, wind protection for large machinery, tank-farm emergency response; the port knowledge base keeps ingesting.

PORTRISKCN/EN/RU

Equipment manufacturing

Process document Q&A, maintenance knowledge, QC report generation.

MFGQA

Human resources

Onboarding training, injury prevention, automated compliance ledgers.

HRCOMPLIANCE

Emergency & first aid

Red Cross emergency knowledge, drill scripts, training and assessment question banks.

RED CROSSDRILL

Industrial park investment

Company profiling for parks, policy matching, investment pitch generation.

PARKLEADS
05 — HOW IT WORKS

The customer asks from a phone
the work runs on our nodes

No technical knowledge and no hardware purchase required. Individuals, small businesses and mid-size enterprises can all order; the heavy lifting is backed by our 11 nodes.

01

Order from a phone

Open WeCom, the mini program or the web page, describe the task, upload the material and get a quote. No software install, no workstation setup.

WeCom / mini program30 s response
02

Platform dispatches

The system splits the task, schedules it and assigns it to an idle node. Progress is pushed to your phone in real time — no phone calls needed.

AUTO SCHEDULINGLIVE PROGRESS
03

Delivered from the nodes

11 nodes produce in parallel and the finished files go straight into the delivery vault. The customer downloads with a code; data stays in our own server room throughout.

190 tok/sNO DATA EGRESS

Three customer types, three ways to buy

From a few dozen yuan per task to an annual licence — pick what you need, no need to go all in.

Individual / studio

From XB-S1 ¥9,800, or pay per task. Submit from a phone, download the result.

PAY PER TASKNO COMMITMENT

Small business

XB-S3 ¥29,800, 5 seats across 3 nodes. Knowledge base plus applications, live within a week.

PACKAGE1-WEEK GO-LIVE

Mid-size enterprise

XB-S10 from ¥69,800, or XB-CLUSTER for full on-premise. Data stays in your own server room.

ON-PREMISEANNUAL
06 — COMPUTE STATION

Four models
first-year service included, renewal from year two

Prices are standard configurations including tax. Custom development, knowledge-base build and system integration are quoted separately.

One control unit is a production line: install the engine, plug in the customer's existing PCs and they can produce images, videos, proposals and ledgers — no need to replace the fleet.

A · Single point starter

XB-S1

¥9,800

1 control unit (engine included) · 1 seat · WeCom entry

For owner-operated shops and solo professionals

YEAR 1 INCLUDEDYEAR 2 ¥3,600/yr
B · Team standard · best seller

XB-S3

¥29,800

1 control unit + 3 existing PCs · 5 seats

Design studios, e-commerce, accounting firms

YEAR 1 INCLUDEDYEAR 2 ¥10,400/yr
C · Enterprise concurrency

XB-S10

from ¥69,800

1 control unit + 10 existing PCs · 20 seats

Factories, chains, mid-size enterprises

YEAR 1 INCLUDEDYEAR 2 ¥18,800/yr
D · Private cluster

XB-CLUSTER

on request

Control cluster · unlimited seats · intranet deployment

Government, state-owned enterprises, classified units

YEAR 1 INCLUDEDYEAR 2 BY SCALE
Standard delivery, not a paid add-on: on-site control-unit deployment · engine pre-installed · integration with the customer's existing PCs · 2 hours of on-site training · remote operations 8×5 in year one · three-year hardware warranty | 5 million cloud tokens included; top up or fall back to the local model automatically at no extra charge; industry knowledge bases are pre-installed as standard, not billed separately
07 — DELIVERY

Three deployment forms
standard products by model, projects quoted separately

Standard requirements go straight to XB-S1 / S3 / S10; work involving custom development or multi-system integration is quoted as a private-deployment project.

AI appliance

XB-S1 / S3: a single control unit with the engine installed, ready on power-up; worker nodes connect to the customer's existing PCs.

1 NODEPLUG & PLAY

Park node

XB-S10 / CLUSTER: one control unit schedules the PCs a park already owns, serving multiple companies inside it.

CLUSTERMULTI-TENANT

Managed operations

Annual licence covering engine upgrades, remote operations and evidence compliance — no in-house AI team required.

MANAGEDSLA
08 — INSIGHTS

Policy and industry notes
updated daily

Policy, industry practice and delivery experience written as short pieces that can be cited by search engines — readable by customers, findable by AI.

2026-11-06

Edge compute for county-level enterprises under the national compute-network plan

What county-level companies actually need is not a hyperscale hub but compute placed beside the plant. (Chinese text)

READ →
2026-11-05

Why “no data leaves the plant” is a hard requirement for enterprise LLMs

Once formulations, customer lists or incident records reach a public cloud they are out of control. On-premise is compliance, not preference. (Chinese text)

READ →
2026-11-04

Confined-space work + LOTO: the six stages a training system must cover

Permit, testing, standby, work, close-out, review. Miss one and the incident review will not stand up. (Chinese text)

READ →

Building edge compute
into regional infrastructure with local governments

The national compute-network plan is led by state-owned and leading enterprises, with local governments contributing sites and incentives. Our role is to bring compute to the end of the industrial chain and hand the AI capability to local enterprises.

No hidden local debt, no reliance on government guarantees, operated under market contracts; we start by activating existing assets and grow into a regional compute service capability.

11nodes in operation
6industry solutions
190tok/s local inference
0data egress

Ways to work together

① Co-build an industrial-park compute node: the government provides the site and incentives, we invest in equipment and operations.

② Industry AI enablement center: on-premise deployment, training and operations for local enterprises.

③ Joint industry models: with trade associations and anchor enterprises, to accumulate regional industrial corpora.

×

Two ways — pick either

Talk to us on WeCom, or place an order yourself in the Xiaobei app.

WeCom · talk to a human
WeCom QR code Scan to add us on WeCom
or call +86 158 8199 1999
Scan with WeChat
Xiaobei app · self-service
Xiaobei app download Xiaobei app
call +86 158 8199 1999 for the installer
Download the Xiaobei app
Or simply call +86 158 8199 1999
CONTENT

Technical articles & delivery cases

Technical articles
Delivery cases
View all articles