In July 2026 we refreshed the Xinference brand and rebuilt xinference.co from scratch. This post covers what changed and why.

Positioning: sovereign AI inference

Xinference is the control plane for AI inference. We orchestrate, observe, and scale model deployments for teams that need to control where their data goes. That has always been true of the platform, and the homepage, the pricing page and this blog all lead with it.

We are an Australian-owned company with data residency in the region you choose. Every deployment path, from the standard bundle to a custom deployment on your own infrastructure, follows one principle: your data never leaves your environment unless you decide it should.

The visual system

The site runs on a rebuilt design system: Open Sans throughout, a single brand blue as the accent color, fully rounded cards and buttons, and no gradients. The look is deliberately plain so that the content stays in front: your models, your metrics, your deployment options.

The site is built around three pillars: sovereignty, cost and performance. On sovereignty, 100% of inference runs in your environment, and your data never leaves. On cost, spend is up to 70% lower than with closed-source APIs. On performance, latency is 2 to 4 times lower, up to 4 times more models are served per GPU, and there are no cold starts. Most platforms make you give up one of these to get another. Xinference is built to deliver all three at once, on your own infrastructure, through one OpenAI-compatible API.

Pricing: one standard bundle

Pricing is built around one standard bundle instead of a tier matrix. The bundle is $10K per month and covers unlimited AI agents, 500 users, and a dedicated private LLM. Additional blocks of 500 users are available as an add-on. Teams that need Xinference to run entirely on their own infrastructure, in the cloud or on-premises, can ask about a custom deployment.

Every path uses the same control plane and the same OpenAI-compatible API, so moving from the standard bundle to a custom deployment does not mean re-architecting your application.

Request routing

The cost savings depend on our routing layer, which was rolling out as of July 2026. It sends each request to the cheapest model that clears your quality bar. Together with KV-cache reuse, continuous batching, and speculative decoding, it is designed to get up to 4 times more useful work out of every GPU, and to let open models carry workloads that often go to a closed-source API out of habit.

The blog

The blog carries engineering write-ups from our founder, tutorials for getting a model running on your own hardware, and updates like this one. The aim is that a technical team evaluating the platform can read a straight answer on a published page instead of waiting for a slide deck.

What hasn't changed

The commitment is no rip-and-replace on day one. Most engagements start as a proof of concept over one to two weeks on the standard bundle, move to a benchmark against your own baseline, and only go to a custom deployment, private cloud, or on-prem once the numbers hold up. The way we work with teams evaluating the platform has not changed with the brand.

Read next: Xinference Model API