Hosted in Australia
Pricing
Most teams start on the Model API and move to dedicated compute as usage grows.
Every option runs the same platform and the same API.
Model API
Frontier open models by the token. Nothing to host.
From $0.30 per 1M tokens
For teams that want to ship on open models today.
Get started →
Cloud deployment
Dedicated GPUs in our cloud, fully managed.
For steady workloads that outgrow per-token billing.
Contact sales →
Private deployment
The full platform inside your own perimeter.
For regulated teams: finance, health, government.
Contact sales →
Xinference + Xagent bundle
Inference plus agents, one stack, one vendor.
For teams building agents on their own models.
Contact sales →Model API
Frontier open models served from Australia, at the same per-token price as the model vendor. Billed per token.
For teams that want to ship on open models today.
Served from dedicated GPUs in Australia.
| Model | What it’s for | USD per 1M tokens (input / output) | |
|---|---|---|---|
 Kimi K3 | Frontier reasoning and agentic workloads | $3.00 in / $15.00 out | Get started → |
 DeepSeek V4.1 Flash | Long-context work and RAG | $0.30 in / $1.20 out | Get started → |
 GLM 5.3 | Agents, reasoning and coding | On request | Enquire → |
| DeepSeek V4.1 Flash costs up to 53% less than DeepSeek on Amazon Bedrock in Sydney. It costs up to 88% less per token than Claude Sonnet 5.5. List prices as of October 2026, USD per 1M tokens. Billed per token. | |||
| Also on request: Qwen, gpt-oss, Llama, Gemma, Mistral, MiniMax, MiMo, Tencent Hy, Nemotron and Step. | |||
| Don’t see the model you need? Talk to us about adding it. | |||
Free guide
Work out the numbers before you scale
The Sovereign AI TCO Guide compares the full cost of running inference yourself with paying for a cloud API, and gives you a four-step payback method to redo with your own figures. 14 pages.
Cloud deployment
Dedicated GPU instances in Xinference Cloud, fully managed, with zero infrastructure to run. Volume discounts available.
For steady workloads that outgrow per-token billing.
| GPU | Memory | Price |
|---|---|---|
| H200 | 141 GiB VRAM | Contact sales |
| H100 | 80 GiB VRAM | Contact sales |
| B200 | 180 GiB VRAM | Contact sales |
| A100 | 80 GiB VRAM | Contact sales |
| RTX6000 Pro | 96 GiB VRAM | Contact sales |
Xinference is NVIDIA optimised yet purpose-built to be hardware agnostic. Enquire with sales if other GPUs are required.
Member of the NVIDIA Inception programXinference on AWS Marketplace
Deploy through the AWS account you already have. Billing counts toward your AWS spend commitments.
Private deployment
Deploy on your own cloud, your data centre, or fully offline environments, with full controls and dedicated support. No data crosses your perimeter.
Xinference + Xagent bundle
Private deployment plus the Xagent agent platform. From $10K a month for 500 users, hosted in Australia.
Community vs Enterprise
Community and Enterprise run the same Apache-2.0 engine. Enterprise hardens it for production: up to 2x throughput, broader hardware and named support.
| Community | Enterprise | |
|---|---|---|
| Engine | Apache-2.0 inference engine | The same engine, plus commercial extensions |
| Hardware support | NVIDIA only (validated) | NVIDIA, AMD, Intel, CPU |
| Scaling | Single cluster | Multi-cluster, multi-region, HA |
| Performance tuning | Default engine settings | Tuned per workload by Xinference engineers, up to 2x throughput |
| Data residency | Manual | Regional deployment, DPA |
| Private offline | Community-supported | Offline Helm chart, dedicated support |
| Support | GitHub community | Named contact, priority response, quarterly engineering reviews |
Pricing questions
How is the Model API billed?
Per token, per model, in USD per 1M tokens. Kimi K3 is $3.00 in and $15.00 out. DeepSeek V4.1 Flash is $0.30 in and $1.20 out.
Can we move between options later?
Yes. Every option runs the same OpenAI-compatible API, so a move is a base URL change, not a rewrite.
Which models can we run?
The Model API serves the models in the table above. On dedicated GPUs or your own infrastructure, Xinference runs 300+ open models, plus your own fine-tunes. Don’t see the model you need? Talk to us about adding it.
Is our data used for training?
No. No training on your data, and no prompt or response content kept by default. They’re processed in Australian data centres or on your own hardware.
What does Enterprise add over Community?
Up to 2x throughput, broader hardware and named support. See the comparison table above.
Can we pay in AUD?
Yes. AUD invoicing is standard for Australian customers.
How do we start?
For the Model API, see how to get started. For dedicated GPUs or a private deployment, ask for a cost assessment and we model your workload side by side.
Last updated