Choose where the work happens.
Run inference in your own infrastructure, a private cloud or an approved hosted service. Define the location, access and logging controls that your data requires.
AiAx OpenWeight
Intelligence.
On your terms.
AiAx offers locally hosted and cloud deployments of open-weight AI, including EU-hosted options. Start with straightforward cloud access, or choose more control over location and retention when your work requires it.
Locally hostedCloudEU hosting availableZero retention where required
Why open-weight
A deployment you choose
Your agents & authorised context
Local or cloud environment
Kimi K3
Qwen
GLM 5.3
DeepSeek V4 Flash
Your chosen model runs here.
Kimi K3
Qwen
GLM 5.3
DeepSeek V4 Flash
Location · Access · Logging · Retention
The model’s trained parameters are available to download and run under its licence. That gives you options beyond a single hosted API. It does not necessarily include the training data or all the freedoms of open-source software.
Why it matters
More choices.
More choices.
Where they count.
Keep your options open.
Separate the model from the company hosting it. Where licensing and technical compatibility allow, move the same weights between providers or retain a tested version.
Adapt to the work you repeat.
Evaluate and fine-tune a suitable model for your terminology, document types or recurring tasks. Compare it against a baseline before putting it to work.
Explore fine-tuning →Match capacity to demand.
Steady workloads can make dedicated infrastructure economical. Compare the full cost of hardware, hosting, operations and evaluation against a managed API. Self-hosting is not automatically cheaper.
Fine-tuning
Teach the model
Teach the model
your way of working.
Fine-tuning adapts a pretrained model using carefully selected examples of the work and the responses you want. It updates model parameters or a smaller adapter, rather than training a new model from scratch.
It is most useful for repeated tasks with a clear standard of quality.
More consistent outputs.
Train on approved examples of classifications, extracted fields and report formats. The aim is fewer corrections on the tasks your team handles every day.
A better fit for your domain.
Teach task-specific terminology, tone and conventions. For example, turn service observations into the categories and action descriptions your team uses.
A smaller model may be enough.
Specialisation can make a smaller model suitable for a narrow workload. Measure quality, latency and total cost before deciding whether it beats a general-purpose model.
How to approach it
Curate approved examples → Tune a compatible model → Test on separate, unseen cases.
Start with a prompting baseline. Methods such as LoRA can reduce training requirements. Deploy only after the tuned version meets your quality checks.
A practical example
A thousand documents.
A thousand documents.
One repeatable workflow.
Consider an agent that classifies incoming service reports and extracts the work to be done. The vocabulary is familiar, the task repeats and the output has a clear structure.
- Start with the data policy.Decide what the model may receive and where that processing may happen.
- Test a suitable open-weight model.Measure extraction quality and exceptions on representative reports.
- Route within those limits.Use the approved endpoint. Escalate difficult cases to an eligible model or a person.
Choose the deployment
Choose the level of control you need.
Some teams need a defined data boundary. Others want a capable model with a simple cloud setup. Your requirements set the level of control.
Locally hosted.
Run models in your own environment. Keep inference close to internal systems and apply your organisation’s access and network controls.
Agree the hardware, operations and support responsibilities for your deployment. Local hosting still needs explicit policies for logs, backups and retention.Cloud, including EU hosting.
Choose straightforward cloud access when you have no special requirements for data location or retention. Focus on model quality, speed and cost. Dedicated hosting can be selected for workloads that need it. Zero retention is configured only where it is a customer requirement.
Standard cloud services use the provider’s data-handling terms. If you need zero retention or a specific region, select an endpoint and configuration that meet those requirements.
Data protection
Support your GDPR requirements.
Choose local hosting or an EU-hosted deployment, with access and retention requirements defined for your workload. Zero retention is provided where the customer requires it and the selected setup supports it.
GDPR compliance also depends on lawful processing, data-processing agreements, security, retention and handling people’s rights. EU hosting alone is not a compliance guarantee.
Confirm the full data path, including model endpoints, backups, support access and subprocessors. Any transfers outside the EEA need an appropriate legal basis for the transfer and applicable safeguards.
Explore security & data handling →A few useful distinctions.
When do you provide zero retention?
We provide zero retention where the customer requires it. It is not the default for every deployment.
Zero-retention inference is an endpoint-specific commitment for handling prompts and outputs. Its scope, exceptions and eligible features must be checked against the provider agreement or the local logging configuration. It is not a blanket promise that no data is ever retained. Workspace history, files and AiAxBrain memory have their own retention rules, so your team can keep the work it needs.
Open weights alone do not guarantee zero retention. Hosting configuration and service terms determine it.
Does it replace frontier models?
It gives you another deployment choice. An open-weight model can itself have frontier-level capabilities. In our routing diagram, “frontier API” describes the separately approved provider route. Compare quality, cost and data handling for each actual endpoint.
Can every open-weight model be used commercially?
Check the exact model licence. Permissions for commercial use, modification, redistribution and hosting vary between releases.
Where does AiAx Router fit?
The Router page explains how task requirements, context and endpoint policy inform model selection. OpenWeight explains why you may want this deployment option in the first place.
Explore AiAx Router →
The next step
Discuss your deployment ↗
Start with your workload.
Bring a recurring task and your data requirements. Use them to compare the deployment options.