Keera Engine
Open models, run efficiently
Keera Engine is the inference engine behind the gateway. It runs any open model on open technology - Kubernetes, vLLM and llm-d - on Swiss GPUs or in your own cluster. More answers per GPU, and no layer is a black box.
What Keera Engine is
An inference engine built from open parts: Kubernetes schedules the GPUs, vLLM and llm-d serve the requests, open models run on top. Above it all sits an ordinary API - whatever talks to OpenAI today talks to Keera tomorrow.
The API
OpenAI-compatible. One base URL, one key, and any client that already speaks to OpenAI speaks to Keera - no new SDK, no plugin.
The models
Model-agnostic: any open model vLLM can serve, plus three Keera fine-tunes and your own. You can read what is inside them and take all of it with you whenever you like.
The inference server
vLLM batches many requests onto one GPU. llm-d sends each one to where its context already sits in cache. The same hardware computes more tokens.
The platform
Kubernetes places the models on the GPUs and scales them with the load. In Swiss data centres or in your own racks, with no internet at all if that is what you need.
The models
The engine runs any open model vLLM can serve - Qwen, Llama, Mistral, Apertus or your own fine-tune. Three fine-tunes of ours run from day one. Pick one per workspace - or let the gateway route: small tasks to the fast model, hard reasoning to the strong one.
| Model | Context | Best for |
|---|---|---|
| Keera Deep | 256k | Agentic work across many files |
| Keera Swiss | 128k | Code and documentation in German, French and Italian |
| Keera Fast | 64k | Completion, tests, commit hygiene |
| Your own fine-tune | - | Internal frameworks and legacy code |
| Any other open model | - | Whatever your use case needs |
Keera Deep and Keera Fast are derived from Qwen-Coder, Keera Swiss from Apertus. All three are Apache-2.0: you can download the weights and take them with you.
Hosted providers - today the OpenAI models - sit behind the same policies, but run on the provider's infrastructure and not on Swiss GPUs. They are blocked by default; a tenant opens them per team and data class, or not at all. Checked September 2026.
Open technology, run by us
Kubernetes, vLLM and llm-d are open source and widely used. We reinvent none of it. We tune the parts to each other and run them - in Swiss data centres, or together with your team in your own cluster.
Every token the engine computes passes through Keera Gateway first: the same policies and the same audit log as everywhere else.
Common questions
What platform and engineering teams ask us most before they try Keera Engine.
Does my data leave the building?
No. The models run on Swiss GPUs or in your own racks; prompts and code go there for inference and come back. They are not stored, the model's answer is not stored, and nothing is trained on either. What is kept is the request itself - who, when, which model - in the log in your own infrastructure.
Are open models good enough?
For everyday work, yes: refactoring across many files, tests, migrations, reviews. On the hardest reasoning the large proprietary models are still ahead - which is why they sit behind the same gateway and can be opened per team and data class. We do not claim open weights win everywhere. We claim you keep the choice.
How does this fit with Keera Gateway?
Keera Engine sits behind the gateway and runs the open models the gateway routes to. The gateway decides who may use which model and writes every request to the audit log; the engine computes the answer. You can start with the gateway and the providers you already have, and add the engine once the models should move in-house.
Try Keera
We put one team on the engine, measure throughput and cost with you and hand you the numbers. After 30 days you decide.