Enterprise
Train it. Run it. Own it.
The whole Run BiOS stack inside your own cloud account or data centre. Train models on data that never leaves, serve them through one OpenAI-compatible API, and keep the weights. We operate the stack. The intelligence it produces belongs to you.
Illustrative. The stack runs where you decide it runs.
Build with leading models
What runs on it
One deployment, the whole platform
Enterprise is not a cut-down edition. Catalog inference, adaptive routing, custom model training and your own endpoints are the same products we run ourselves, installed where you decide they run.
Custom model training
Expert models, domain-specific models and vision tuning, trained on your own data inside your own environment. This is the centre of the enterprise product.
What you can trainCustom model serving
Your finished checkpoints on dedicated endpoints inside your environment, called by model id like any other model.
How serving worksCatalog inference
Every model in the served catalog, answering from inside your deployment behind the same OpenAI-compatible API, for the work a custom model is not needed for.
See the catalogAdaptive routing
Available in the deployment when a workload spans several models: one endpoint that reads each request and sends it to the right model for the task.
How Adaptive worksAgent & framework integrations
LiteLLM, LangChain, CrewAI, coding and messaging agents, all pointed at your endpoint with the same three values.
See integrationsThe boundary
What sits inside your walls, and what crosses them.
One line crosses the boundary, and it carries software inwards. Your datasets, your traces, your checkpoints and the endpoint serving them never leave the network they were created on.
Your cloud account or your data centre. Your network policy, your identity, your storage.
- Signed images
- Engine releases
- Licensing
- Support and on call
No customer data, ever. The only line that crosses the boundary carries software in.
What you get
Enterprise control without rebuilding the platform
Four things, and none of them is a compromise: your data stays yours, your developers keep one API, the operations are our job, and the weights you train are yours.
Custom model training
Train expert models on data that never leaves
The whole training platform runs inside your deployment: your own corpus becomes a model that knows your domain, and the checkpoint it produces is a file you own outright.
Expert models
A model that does one job better than any general-purpose model. Fine-tune on your own task examples, then align it, until the format, the tone, and the judgment hold on every call instead of most of them.
Domain-specific models
Continued pre-training on your own corpus of contracts, policies, tickets, filings and code puts your vocabulary and conventions in the weights, so the model stops needing a prompt to explain your business to it.
Vision-language models
Tune a VLM on the documents and images your workflow actually handles: forms, layouts, diagrams and inspection frames, the ones a general model was never trained on.
The Run BiOS training platform
Your data · Your GPUs · Exported checkpointsPre-training, continued pre-training and supervised fine-tuning, for language and vision language models, as LoRA or QLoRA adapters, or a full fine-tune of every weight. Start with the cheap method, escalate to the thorough one when the evals say it plateaued, all from the same job config.
Custom model inferencing
Serve the model you trained like any other
A checkpoint is only worth what you can run. Your custom models serve from dedicated endpoints inside your environment, on the same API your applications already speak.
One API, one new model id
Your fine-tuned model answers on the same OpenAI-compatible endpoint as the catalog. Change the model id and keep the client, the SDK, and the code path you already shipped.
Dedicated, isolated capacity
A custom model serves from its own dedicated endpoint on capacity inside your environment, not shared with anyone else, sized to the workload it serves.
Catalog and custom, side by side
The models you trained and the models we serve sit behind the same API surface, so one application can call both: the specialist for the work that needs it, the catalog for everything else.
Illustrative. One POST /v1/chat/completions, two model ids.
BiOS Conscious Loop
And then it starts getting better on its own.
This is the part that only works inside your own walls. Every interaction your agents and applications serve leaves a record, and the records never leave the building. On a schedule you set, the system turns that work into training material, trains the next version, tests it against the one in production, and promotes it only if it wins.
We call the mechanism the BiOS Conscious Loop, and what it leaves behind is Specialised Intelligence: a model that is quietly unbeatable at the one thing your company has done ten thousand times, sitting on your storage, under your name.
Note where the improvement lands. Retrieval and long context put knowledge in front of the model at the moment it answers, which is the right tool for anything that changes daily and is worth having. This works on the model itself: the weights move, so the judgement is there by default, on every call, without being re-supplied.
You can design that loop yourself, down to which workflows feed it and what counts as an outcome worth learning from. Or Run BiOS reads your traces and proposes the whole thing for you to approve or change. Your decision either way, and the intelligence it produces is yours either way.
How the BiOS Conscious Loop works- 01Serve. Your applications answer real work through Run BiOS.
- 02Record. Every interaction and correction is kept, on your storage.
- 03Curate. Good outcomes become examples, corrections become lessons.
- 04Train. A new version, on settings you approved once.
- 05Evaluate. Scored against production, on tests you wrote.
- 06Promote. Only if it wins. Otherwise nothing changes.
You set the cadence and the promotion gate. Every version is kept, so going back is one action.
Built for
When renting inference stops working
Most teams should use the managed platform. These are the situations where running the stack yourself is the only version that works.
Data that cannot leave
A workload with rules about where it may be processed, and a boundary it is not allowed to cross.
GPU capacity you already own
Hardware already bought or committed, and no modern serving platform running on top of it.
A private environment
Clusters a public endpoint cannot reach: isolated networks, private links, or fully disconnected sites.
Models of your own
A corpus or a set of task examples that should become your model, not a prompt pasted into someone else’s.
One platform, not five vendors
Catalog inference, training and your own endpoints behind a single API instead of a vendor per capability.
What does deployment look like?
01
Bring your capacity
GPUs you already own, or capacity you procure in your preferred cloud or data centre. The cluster stays yours: your account, your nodes, your network policy.
02
We install the stack
Run BiOS deploys the training and serving stack into your cluster with a dedicated, revocable setup credential, then scopes every runtime component down to its own identity.
03
Validate together
GPU readiness, endpoint reachability, routing, and autoscaling are checked with your team before production traffic touches the cluster.
04
Train, serve, and keep improving
Your applications call the same OpenAI-compatible API they already use. Training and inference both run inside your boundary, we operate the stack day to day, and the BiOS Conscious Loop can be switched on whenever you are ready for it.
Questions teams ask before the first call
Where does training and inference actually run?+
Inside your environment: your cloud account or your data centre. The stack is deployed into Kubernetes infrastructure you own, and request and response handling, fine-tuning runs, and the traces the BiOS Conscious Loop learns from all stay within your network boundary.
Is it the same API as the cloud platform?+
Yes. One OpenAI-compatible API, the same model catalog, and BiOS Adaptive routing across your on-prem deployment and our managed capacity. Code written against one runs against the other.
Who operates the deployment?+
Run BiOS operates the serving stack: model deployment lifecycle, upgrades, GPU health monitoring, autoscaling, and observability. Your team owns the environment, the hardware, and the network policy.
What about fine-tuning?+
Fine-tuning runs on dedicated GPUs inside your environment, and the checkpoint is a file you own. Export it and serve it inside your deployment or anywhere else. Our own training engine is written from the ground up, down to the model implementations, and covers pre-training, continued pre-training and supervised fine-tuning, as a full fine-tune or as an adapter, for language models and vision language models.
Does the model actually get better over time, or is that marketing?+
It is a mechanism, and you control it. Once the loop is switched on, the work your teams already do becomes the training material for the next version, on a cadence you choose. A new version is only promoted if it beats the one already serving, on tests you wrote. Every version is kept, so rolling back is one action. The full explanation is on the BiOS Conscious Loop page.
Who designs the loop, us or you?+
Whichever you prefer. You can specify all of it: which workflows feed the loop, what counts as an outcome worth learning from, the tests a new version has to beat, the cadence and the sign-off. Or Run BiOS reads your traces and proposes that plan for you to approve, change or reject. Nothing runs until you have said yes, and the model that comes out is yours in both cases.
If we stop working with you, what happens to the models?+
They keep working, because they were always yours. The weights, the checkpoints, the traces and the training data stay where they have always been, on your storage. There is nothing for us to take back.
What if traffic outgrows our cluster?+
Hybrid capacity is an option: eligible workloads can overflow onto Run BiOS-managed capacity during spikes or while your own GPUs are being expanded or remediated. Which workloads may overflow, when, and how routing behaves are agreed with your team up front.
What is the difference between an expert model and a domain-specific model?+
An expert model is trained on task examples with supervised fine-tuning, so it holds a format, a tone and a judgment call reliably. A domain-specific model is trained on a corpus with continued pre-training, so your vocabulary and conventions live in the weights. Most production models use both: pre-train on the corpus, then fine-tune on the task.
Can we serve our own fine-tuned models inside the deployment?+
Yes. A finished checkpoint deploys to a dedicated endpoint in your environment, behind the same OpenAI-compatible API and called by model id like any catalog model. You can also export the file and serve it somewhere else entirely.
What does it take to get started?+
A Kubernetes cluster with NVIDIA GPU nodes, network access that lets us manage the stack, and a conversation with our team about your environment. We confirm fit, then install and validate with you.
Proof
Workloads already running this way
Anonymized engagements, as enterprise work should be. The document-heavy ones, where owning the model mattered as much as running it.
Legal
Legal: diligence at deal speed, with a model the firm owns
Contract review was bottlenecked by hours and a rented general model. A fine-tuned model on the firm’s own clause library changed that.
Drafts arrive in the firm’s own clause languageRead the studyRetail & e-commerce
Retail: a storefront that writes and answers in its own voice
Product copy and support answers ran on a premium model across the whole catalog. Right-sizing the models changed the economics.
The bill stopped tracking the sales calendarRead the studyManufacturing & logistics
Manufacturing: manuals that answer back
Maintenance knowledge lived in binders and inboxes. A fine-tuned model over the company’s own documentation brought downtime down significantly.
Downtime brought down significantlyRead the studyBring us your environment.
Tell us where training and inference need to run, whether that is a cloud account, a data centre or GPU capacity you already have, and we will confirm fit, install, and validate with your team.
One OpenAI-compatible API · BiOS Adaptive routing · We operate the stack