Frontier open models,
hosted in Australia.
GLM, DeepSeek and Gemma behind one OpenAI-compatible API, priced below big-cloud rates for the same models. Switching takes two lines.
client = OpenAI(
base_url="https://api.xinference.co/v1",
api_key="XINFERENCE_API_KEY",
)
# requests stay in Australia.The platform behind these production deployments:
Models
| Model | What it’s for | USD per 1M tokens |
|---|---|---|
 GLM 5.3 | Agents, reasoning and coding | Contact sales |
 DeepSeek 4.1 Flash | Long-context work and RAG | Contact sales |
 Kimi 3 | Frontier reasoning and agentic workloads | Contact sales |
| Xinference serves 300+ open models, plus your own fine-tunes on the same endpoint. Contact sales for anything you don’t see. | ||
Pricing logic
Why it costs less
Open models cost less than closed ones, and our rates are lower again.
Open models already undercut closed models
On Bedrock’s own list pricing, DeepSeek runs at $1.35 in / $5.40 out per 1M tokens. Claude Opus sits at $5 / $25. Same tasks, a fraction of the bill.
We price the same open models below Bedrock
They run on our own infrastructure in Australian data centres, on our optimised proprietary serving stack. High performance, low cost, and a better price than Bedrock for the same models.
Prove it on your workload
Send us your current usage or bill. An engineer builds the side-by-side cost model with you, before you commit.
Get a cost assessment→Most production workloads already run on open models, with the same or similar results as closed APIs.
Bedrock figures are public AWS list pricing, September 2026, USD per 1M tokens.
The details engineers ask about
OpenAI-compatible endpoints
Drop-in /v1/chat/completions and /v1/embeddings. Existing OpenAI SDK clients work unchanged.
SSE token streaming
Set "stream": true and tokens arrive as server-sent events.
TTFT and TPOT metrics
Time to first token and time per output token, measured on every request. Teams see 2 to 4x faster inference after tuning.
Per-request logging
Every call inspectable: tokens in, tokens out, latency, model version.
Secrets management
Scoped API keys, so credentials never live in your codebase.
Sovereign-first hosting
Hosted in Australian data centres, under Australian jurisdiction, with vLLM and SGLang engines behind one API.
curl https://api.xinference.co/v1/chat/completions \
-H "Authorization: Bearer $XINFERENCE_API_KEY" \
-d '{ "model": "gemma-4", "stream": true,
"messages": [{ "role": "user", "content": "hello" }] }'
# tokens stream back over SSE. TTFT and TPOT logged per request.Sovereignty
Inference stays onshore
An offshore provider storing your data in Australia still runs the model overseas. Here, the models and the inference stay onshore.
Data stays in Australia
Prompts and outputs are processed in Australian data centres, or on your own hardware.
Your data stays yours
Your data is never used to train or improve models, ours or anyone else’s.
Processed, then discarded
Requests are processed and discarded, not logged for review or analytics.
Models you can take with you
Open models behind an API you already know. Leave any time and keep the weights.
Getting started
Start with a PoC
Most enterprise AI projects start as a PoC. Call the API today, put a week of real traffic through it, and set the bill next to your current one. An engineer runs the comparison with you.
Two-line switch, first tokens served.
Your workload, side by side with your current stack.
Compare the numbers side by side.
As your usage grows, the same API moves from tokens to dedicated GPUs to your own infrastructure.
See deployment options→Questions
How is the Model API billed?
Per token, per model. Contact sales for current rates and volume pricing.
Can we move to dedicated compute later?
Yes. Every option runs the same OpenAI-compatible API, so a move is a base URL change, not a rewrite.
Is our data used for training?
No. No training on your data, no retention by default, processed in Australian data centres or on your own hardware.
Which models can we run?
300+ open models including GLM, DeepSeek, Gemma, Qwen and Llama, plus your own fine-tunes.
Do you support function calling and structured output?
Yes. Function calling, structured output (JSON mode), streaming, embeddings and vision models are all served through the same API.
What are the rate limits?
Default limits apply per model. Higher concurrency is available on request through sales.
Is there an SLA?
Yes. Paid plans come with an uptime SLA and defined support response times. Ask sales for the current terms.