kawkav
Sign in
Kawkav / Infrastructure / Sovereign AI

AI · private models

Your models. Your data. Your infrastructure.

Most AI runs on someone else's cloud, on someone else's terms. Kawkav puts private LLMs inside your own tenant — under Swiss and EU jurisdiction, on open formats you can take with you. Your prompts and documents never leave your boundary and never train anyone else's model.

12
Models on the mesh
0
Egress — ever
CH · EU
Data residency
€9.90
Swiss S3 / mo
Sovereignty by design

AI without surrendering control.

Four principles hold across every product we ship. They are not marketing — they are how the platform is built.

Private LLMs, your tenant

Models run inside your dedicated tenancy — not a shared public endpoint. Prompts and documents never leave your boundary and never train anyone else's model.

Swiss & EU residency

Data stays in Swiss and EU regions by default. GDPR is the starting point, not an add-on — export, deletion, DPAs and a clear sub-processor list come standard.

Your keys, no egress

Bring HSM-backed keys for encryption and audit-log signing. Egress is blocked at the network level, not just by policy — inference stays on sovereign ground.

No lock-in, open by default

Open-weight models, open formats, open APIs. Take your weights, your vectors and your data and leave whenever you want. We earn the renewal — we don't trap it.

One sovereign mesh

A single Kawkav mesh powers everything — chat, voice, vision, multilingual. Compact 8B/7B, a balanced 14B tier, a flagship large model, an Arabic-native model and a vision model, served as one fabric.

Audit-grade logs

Every request logged with model version, parameters, latency and token counts. Append-only, signed, and exportable as CSV or streamed straight to your SIEM.

Model catalogue

Open-weight LLMs, closed-tenant inference.

Curated, quantized and packaged for cost-efficient serving on your tenant. OpenAI-compatible endpoint, per-token billing, reserved capacity on demand.

llama-3.3-70b
Meta · open weight · 128k ctx
Top-tier generalist. Strong reasoning, multilingual. Great default for new deployments.
€0.40 / 1M tok
mistral-large
Mistral · open weight · 128k ctx
European frontier model. Excellent multilingual reasoning, code and structured output.
€0.55 / 1M tok
qwen-2.5-72b
Alibaba · open weight · 32k ctx
Strong coder in Python and SQL. Best-in-class for tool calling and agentic workflows.
€0.38 / 1M tok
deepseek-v3
DeepSeek · 671B MoE · 128k ctx
Mixture-of-experts at the frontier, ~37B active per token. Premier reasoning and math.
€1.20 / 1M tok
phi-4
Microsoft · open weight · 14B
Small but mighty. Reasoning that punches above its weight — ideal for cost-sensitive work.
€0.12 / 1M tok
gemma-3-27b
Google · multimodal · 128k ctx
Best-in-class multimodal. Reads images, OCRs documents. Strong for visual-document flows.
€0.22 / 1M tok
llama-3.2-3b
Meta · edge · sub-100ms TTFT
Edge-friendly mini-model. Perfect for autocomplete, classification and intent detection.
€0.04 / 1M tok
nomic-embed-v2
Nomic · embeddings · 768 dim
Long-context multilingual embeddings, 8192-token chunks. Pair with your tenant Qdrant or Chroma.
€0.02 / 1M tok
+ custom fine-tune
Your weights · LoRA / QLoRA
Bring your own LoRA, QLoRA or full fine-tuned weights. We host, version and serve. SafeTensors only.
€89 / mo hosted

Your existing OpenAI code works — change two lines, the base URL and the key. Same endpoints, same SDKs. Need dedicated silicon? GPU on-demand →

Plans

Start private for €9.90. Reserve a GPU when you scale.

Per-token, monthly minimum, or fully reserved single-tenant silicon. EU data residency on every tier — no co-tenancy on the private plans.

Most popular
Sovereign Swiss S3
€9.90/mo
  • Private LLM endpoint on your tenant
  • Swiss / EU data residency
  • OpenAI-compatible API key
  • All catalogue models, per-token
  • Audit log + GDPR export
Deploy Swiss S3 →
Tenant inference
€89/mo
  • Dedicated tenant, no co-tenancy
  • + €0.30 / 1M tok · 70B class
  • 1M tokens/day included
  • Custom fine-tunes hosted
  • Customer-managed keys (HSM)
  • 99.9% SLA + SIEM stream
Start tenant →
Reserved capacity
€1,200/mo
  • Dedicated GPU — A100 · H100 · B200
  • Unlimited tokens at served capacity
  • Pin specific model versions
  • Air-gap / on-prem option
  • DPA + SCC + DPIA support
  • 99.99% SLA available
Reserve capacity →
Questions

The honest answers.

Where does inference physically happen?

On dedicated GPUs in Swiss and EU regions under our own control — no US hyperscaler underneath. You pick the jurisdiction at deploy, and your prompts, completions and embeddings never leave the tenant.

Do my prompts or documents train your models?

Never. Models run inside your own tenancy, not a shared public endpoint. Egress is blocked at the network level, not just by policy, so your data cannot be used to train anyone else's model.

Is there vendor lock-in?

No. Every model is open-weight, served on open formats through an OpenAI-compatible API. Export your weights, vectors and data and leave whenever you want — we earn the renewal.

Can I bring my own fine-tuned model?

Yes. Upload a LoRA, QLoRA or full fine-tune (SafeTensors), and we host, version and serve it on your tenant from €89/mo.

How do I pay and provision?

Deploy from your Kawkav console — the wizard provisions the endpoint live and streams progress. Billing is per-token, capped by your plan; no third-party checkout.

Ready when you are

Build on sovereign AI.

Your models, your data, your tenant — inside Swiss and EU jurisdiction, with no lock-in. Private inference is one click from live, from €9.90 a month.

Pairs well with