tandemcompute: run open-source AI on Irish infrastructure
models

Run open-source AI on Irish infrastructure.

Access leading open models through tandemcompute APIs without managing the GPUs underneath.

tandemcompute hosts selected open-weight models on its Irish RTX fleet and exposes them through OpenAI-compatible endpoints, with EU-based data processing.

sovereign model api

Open models, kept in Ireland.

An OpenAI-compatible API running the best open models on our own GPUs in Ireland. Your data never leaves the country and is never stored.

inference.pyopenai-compatible
from openai import OpenAI

# Point your existing client at tandemcompute.
# Inference in Ireland, zero data retention.
client = OpenAI(
    base_url=TANDEMCOMPUTE_API_URL,  # issued with your API key
    api_key="YOUR_API_KEY",
)

reply = client.chat.completions.create(
    model="qwen/qwen3.6-35b-a3b",
    messages=[{"role": "user", "content": "Summarise this contract."}],
)

Data stays in Ireland

Inference runs on tandemcompute hardware in an Irish Tier III facility.

Zero data retention

Prompts and outputs are never stored, and never used for training.

Drop-in OpenAI compatibility

Point your existing OpenAI client at tandemcompute. One line of code changes.

A fraction of the cost

Well below closed frontier APIs, with the right model for each job.

Committed volumes

Monthly token commitments, or dedicated endpoints at contract rates.

One supplier

The same account covers bare metal, private AI and the API.

Cutting-edge open models, hosted in Ireland.

Three families: GLM, Kimi and Qwen, at their current releases. Exact versions are confirmed at launch.

providerZ.aiGLM 5.3753B1MDedicated, whole nodeFlagship reasoning, coding and agents. FP4 on one node.Z.aiGLM 5.3 Flash320B1MShared and dedicatedFast multimodal GLM for high-volume traffic.MoonshotKimi K32.8T1MDedicated, whole nodeFrontier agentic and long-context model. FP4 on one node.AlibabaQwen3.8-Max2.4T1MDedicated, whole nodeAlibaba flagship, open text weights. FP4 on one node.AlibabaQwen3.8-27B27B262KShared and dedicatedDense vision-language model, fits one GPU.AlibabaQwen3.6-35B-A3B35B262KShared and dedicatedEfficient general model, fits one GPU.OpenAIgpt-oss-120b117B131KShared and dedicatedOpenAI open-weight model for cheap high-volume work.

browse the full model library →

Versions as of September 2026; the table tracks each family’s current release. An 8-GPU tandem node carries 768 GB of memory: models above ~350B run in FP4 on a dedicated node; everything else fits in FP8 with room for context.

One OpenAI-compatible API. Shared, dedicated or private.

shared

Shared endpoint

Access to the model table through a fully managed API. Change the base URL and go.

best for
  • Early-stage products
  • Low or variable traffic
  • Trying an EU-hosted model in an afternoon
request an api key →
dedicated

Dedicated endpoint

A model on GPUs reserved for you, tuned for throughput or responsiveness, with no shared rate limits.

best for
  • Production traffic
  • Latency-sensitive applications
  • Throughput you can plan around
request a dedicated endpoint →
tailored

Private AI & managed workloads

Your own or a fine-tuned model, benchmarked and deployed by tandem to your requirements and SLA.

best for
  • Proprietary or fine-tuned models
  • Bespoke optimisation
  • Contracted performance
request benchmarking →
at a glance
who it’s for

Developers and companies that want an endpoint, not a cluster.

setup time

API access on request as capacity comes online; days for a dedicated endpoint.

data residency

Processed in Ireland on tandem-controlled infrastructure, EU jurisdiction.

support

A named tandem engineer, by email and phone, Irish hours with out-of-hours cover.

The questions that usually come first.

Can I run my own model?

Yes, through a private AI & managed workloads. Send the weights or a Hugging Face reference and we benchmark it on our nodes before you commit.

Is the API OpenAI-compatible?

Yes. Existing clients work by changing the base URL and key. Streaming, tool calls and JSON mode are supported.

Where is my data processed, and is it kept?

Requests are processed in Ireland on tandem-controlled infrastructure. Prompts and completions are not stored beyond the request unless you enable storage or logging. Nothing is used for training.

What about rate limits?

Shared endpoints carry per-key limits published with the key. Dedicated endpoints have none beyond the capacity you reserve.

Can I have a dedicated instance?

Yes. A dedicated endpoint runs on GPUs reserved for you, from one GPU to a whole node, billed monthly.

Other ways to work with tandemcompute.