(1)

Looking back at the almost past year of 2025

It is a year for domestic large models like Qwen, DeepSeek

to cross the turning point

Beyond models, engineering technology

is also accelerating its evolution and implementation

The AI gateway we are working on

has transformed from a concept that sounds a bit new

into essential infrastructure that many clients

cannot do without when implementing AI

(2)

I remember at the beginning of the year when DeepSeek R1 was released

we held an urgent meeting overnight

to assist enterprises in quickly

transitioning from OpenAI to domestic large models

A few days later, Qwen2.5-Max was released

we responded quickly again

not to seize the spotlight

but to let those brave AI companies know

we are moving forward alongside them

(3)

At that time, we were also considering

that just supporting model switching is not enough

for large models to be truly useful

they need to be networked, have permission controls

to prevent abuse and allow for effect observation, etc.

Thus, in plugin form

we were the first in the industry to support networked searches

to help large models access full-text searches

and we laid out our practical experience

serving early clients

summarizing

eight typical scenarios of the AI gateway

To be honest

at that time we didn't expect

this would become the first complete AI gateway capability map in the industry

But for us

it is a map of user pain points

every piece was honed from real business

(4)

On the path of open source

we continuously expand the capability range of the AI gateway

for example, we opened up

a load balancing algorithm optimized for large models

cutting down the initial token latency by 50%

(5)

We open-sourced the subproject HiMarket

based on Alibaba's internal

AI open platform IdeaLAB

allowing every enterprise to have

a dedicated

model, application, and interface unified management platform

which facilitates architects and operations personnel

and also makes it easier

for programmers, operations, designers, and other

users of AI tools

Adding AgentScope and AgentRun

provides a relatively complete AI development toolchain

(6)

During the peak of MCP

many were troubled by the conversion of existing APIs to MCP

we open-sourced a low-code conversion tool

and built an MCP marketplace

to help developers

quickly connect to over 50 high-quality MCPs

(7)

A month later

Higress was selected

for the MCPMarket MCP Server

Global Top 100 Ranking

Major open-source client agents

began integrating Higress

Alibaba's Taotian business also used Higress

to quickly convert internal HSF services

into MCP Servers

This was again used by a corporate client after

Tongyi Qianwen, Bailian, PAI, Gaode, and Ele.me

(8)

This year

we also served many external clients

Ctrip shared at the China Trusted Cloud Conference

how they solved real challenges in launching large models through

Higress

becoming a reference example for many clients implementing the AI gateway

The Ant Group SOFA team

released SOFA Higress based on Higress

seeing their open-source project

deeply adopted by financial-grade clients

is a strong boost for us to continuously lead the AI gateway

Also, our angel user Sealos

shared on Reddit

their experience migrating from Nginx Ingress to Higress

and their complete journey of nearly a hundredfold performance improvement

has garnered praise from foreign developers

Higress is being recognized by the world

(9)

In addition to serving corporate clients

we have always valued relationships with developers

interacting with developers in various ways

for example, participating in

KubeCon, Zhejiang University Taiyi platform, and the Summer of Open Source at the Academy of Sciences

With the support of the Open Atom Foundation

we held the first Higress AI Gateway Developer Challenge

11 teams entered the finals

competing in technology and creativity around AI Agent, RAG, and intelligent routing

(10)

However, not every open-source project develops

smoothly

In November, we saw a very regrettable news

Ingress NGINX announced its retirement

The K8s community listed Higress and other open-source projects

and their cloud products

as officially recommended alternatives

While comforting, we also reflected

that open source cannot rely solely on enthusiasm to generate electricity

it must have sustainable commercial support

to go further

(11)

We launched

Serverless instances of Higress Enterprise Edition

for less than one-tenth of the resource cost

you can set up a high-availability AI gateway

Clients told us

that AI engineering

really can be this simple

For example, Semir

achieved unified management of multiple models and MCPs

through Higress Enterprise Edition

with overall efficiency improvement of 30%

quickly becoming a benchmark for AI implementation in the industry

(12)

Honor comes from accumulation

winning the

Outstanding Open Source Community Award

at the Wuzhen World Internet Conference

Community contributor Xing Guofu won

the Most Valuable Contributor Award

The AI gateway based on Feitian Enterprise Edition

won one of the three major innovation practice awards

and also participated in drafting the industry standards for AI gateways

led by the China Academy of Information and Communications Technology

with client cases from Ctrip, Guotai Junan Property Insurance, and Junrun Smart

being selected as best practices

(13)

These recognitions from the industry

will always be cherished memories

Looking back at this year

there has not been a single month that was easy

but every step was taken on solid ground

Writing code, doing open source, taking on demands, and running client projects

we firmly believe

open source

is our commitment to the industry

and business

is our responsibility to long-termism

In the future

we hope every enterprise that wants to use AI

can steadily take their first step

👣

👣