Infinigence AI (Wuwen Xinqiong)

What Infinigence AI is

Infinigence AI, known in Chinese as Wuwen Xinqiong, presents itself as a general artificial intelligence infrastructure provider^[1]. It does not sell a chatbot or a single model; it sells the layer underneath them. The company frames its core business as a middle layer between M kinds of models and N kinds of chips — software that lets many different models run on many different kinds of hardware^[1].

Platform and products

Its main product is an Agentic Infra platform that covers large-scale model training and inference for AI-native companies, meaning companies whose products are built around models rather than around conventional software^[1]. Training is the expensive process of fitting a model's parameters to data; inference is the cheaper but far more frequent process of running the finished model to answer requests. A platform that handles both has to manage compute, memory and data movement across a whole cluster, not just a single machine.

A separate product line works on the hardware side: it manages chips of different architectures and schedules clusters of ten thousand accelerators^[1]. Clusters at that scale mix chips from several vendors and generations, and scheduling means deciding which job runs on which hardware at which time, and keeping the whole cluster busy.

Above that sits Agentic MaaS, a service the company describes as a token factory that sells model capacity to enterprises and developers^[1]. MaaS, model as a service, means customers call a model through an interface instead of hosting it themselves; the unit of billing is the token, the small chunk of text a model reads or writes.

The company also markets an inference acceleration engine called InfiniMizar, which adapts many models to many hardware backends^[1]. Moving a model onto new silicon usually means rewriting kernels, the low-level routines for the operations a network performs. An engine that hides this difference is what makes the M-models-by-N-chips idea practical rather than a slogan.

Smart-hardware solutions built on InfiniMizar are offered at module level and as desktop devices^[1], which suggests the same engine is meant to run outside the data centre as well as inside it.

Three core technologies

The company names three core technologies: multi-heterogeneous scheduling, hardware-software co-design, and agentic AI for infrastructure^[1]. The first is about placing work on a mix of different hardware. The second means designing software and hardware together so that each fits the other, instead of treating the chip as a fixed black box. The third applies AI agents — software that plans and takes actions on its own — to running the infrastructure itself.

Megrez models

Infinigence AI also releases models of its own. Megrez-3B-Instruct was trained entirely in-house^[2]. The Megrez2 technical report lists authors from Infinigence AI, Shanghai Jiao Tong University, Tsinghua University and Aalto University^[3], so part of the model work is a collaboration with academic groups. Its GitHub organisation publishes the Infini-Megrez code and model weights together^[4], and the weights are released under Apache-2.0^[2][4], a permissive licence that allows commercial use and modification. That matters to companies that want to try or fine-tune a model without negotiating a licence first.

Where it sits

The company says it serves more than one hundred leading customers across industries^[1]. Its product list runs from cluster scheduling through a serving engine to end devices, and it is one of the few organisations in this space that ships open model weights alongside the infrastructure software, so the Megrez releases give outsiders something concrete to inspect when judging the stack.