Security & Data ResidencyData Sovereignty

    Data sovereignty, built in - not bolted on.

    Kimchi gives you three ways to run models - and complete transparency about where your data goes in each one. You choose your sovereignty level per workload, instead of an all-or-nothing decision.

    kimchi - data residency
    ~/project · main
    kimchi models --show-residency
    MODEL
    MODE
    RUNS ON
    LEAVES
    glm-5-fp8ServerlessKimchi · FR/IL/US✗ never
    minimax-m3Self-Hostedyour k8s cluster✗ never
    claude-opus-4BYOKAnthropic (your key)✓ your key
    residency verified · 2 of 3 workloads never leave
    Three Modes - You Choose

    Pick a sovereignty level for every workload.

    01 · Serverless

    Data never leaves the Kimchi platform

    Hosted by us, on our own GPUs.

    • Runs on Kimchi-owned GPUs in France, Israel & the US
    • Pretrained open-weight models only
    • Nothing stored, no prompt ever reused
    • No third-party or model-provider infrastructure
    DATA LEAVES · NEVER
    FR · IL · US
    02 · Self-Hosted

    Data never leaves your environment

    Maximum control, zero trust required.

    • Same models, deployed in your own Kubernetes cluster
    • Zero dependency on Kimchi infrastructure
    • You control the exact physical processing location
    • No trust extended to any third party - including us
    DATA LEAVES · NEVER
    YOUR INFRASTRUCTURE
    03 · Frontier · BYOK

    Your key, your explicit data flow

    The one mode where data leaves - by your choice.

    • Connect OpenAI, Anthropic, Mistral & others
    • Traffic goes directly to that provider, under their terms
    • Kimchi doesn't intercept, store or alter it
    • Surfaced clearly - never any ambiguity about the mode
    DATA LEAVES · YOUR KEY
    DIRECT TO PROVIDER
    Where the Models Come From

    Model origin isn't data destination.

    Model origin
    Data destination

    Several of the strongest open-weight coding models come from Chinese labs - DeepSeek, Zhipu/GLM, Moonshot/Kimi, MiniMax. Open-weight means the weights are published and run entirely independently of their origin. On Serverless & Self-Hosted the model runs on our GPUs or yours - never the lab's. There is no network path from your data to DeepSeek, Zhipu, Moonshot or MiniMax. And we never train: no prompt, code snippet or output is ever used to tune any model.

    Open-weight Labs
    Open Source LLMs
    Weights published openly
    downloaded once
    Runtime
    Kimchi GPUs · your cluster
    The model runs here
    your prompts & code
    You
    Your data
    Prompts, code, output
    ✕ No return path.Nothing flows from the runtime back to DeepSeek, Zhipu, Moonshot or MiniMax - it never phones home.
    Across All Three Modes

    No training, ever. Kimchi is infrastructure, not a model trainer - pretrained models only.

    Enterprise-grade by architecture - enforced structurally, not by a policy promise.

    No data stored, no prompt reuse on Kimchi-operated infrastructure (Serverless & Self-Hosted).

    Full auditability. You choose the mode, so the data flow is never ambiguous.

    Sovereignty means control  over where your data lives.

    Need country-specific processing guarantees? Self-Hosted gives you that directly. Talk to us about the right mode for your workloads.