Platform · Models

KERNSTACK Inference

Inference on KERNSTACK comes in two honest forms. Catalogue models are callable by name from project code — text-embed-v1 for embeddings at 3 credits per 1,000 tokens, sentiment-multilingual-v1 at 2 credits per 1,000 tokens, vision models at 3 credits per image — with no vendor API key in your secrets. Custom inference endpoints declare your own model (runtime, optional custom image, model bucket and prefix) and meter their compute.

Model Inference

Talk to us Read Inference Docs → Software Lifecycle Cloud →

Why teams replace assembled cloud services with KERNSTACK Inference

Teams otherwise run an external model API and a second bill, with keys copied into every environment and usage attributed by archaeology. On KERNSTACK the model call is a project resource: named, metered to the same balance, and visible beside the code that spends it.

Catalogue: intelligence without procurement

Embedding a corpus, classifying tone in many languages, scoring an image — each is a named call at a posted credit price. Workflow steps can invoke models directly (an ai step carries model, prompt, context from earlier steps, and an output schema), so a document pipeline’s “extract the total” is a declared step rather than glue code around somebody’s SDK.

Custom endpoints: your model, platform plumbing

Declare an inference_endpoint when the catalogue does not carry the model: a fine-tune, a specialist classifier, a model under your own licence. It runs with its declared runtime and resources, reads its weights from your bucket, and callers treat it exactly like a catalogue model — the swap is a call-site change in one project, not a re-architecture.

No API keys for catalogue models; usage attributes to the calling function, worker, or workflow step
Vision at 3 credits per image: a 10k-upload month is a ≈30,000-credit line, statable in advance
Workflow ai steps accept context references from earlier steps and output schemas for structured results

SDKs & Language Integration

Every KERNSTACK environment vendors typed SDK packages into .kernstack/ so both engineers and coding agents in Agent Sandboxes can call Inference without external credentials:

Go SDK: github.com/coding-berlin/kernstack-sdk-go/sdk_inference
Node.js SDK: kernstack-sdk
Python SDK: kernstack_sdk
PHP SDK: kernstack/sdk (Composer)

European alternatives to the hyperscaler equivalent

Amazon SageMaker →

Implementation guides for Inference

Go + Inference → Python + Inference →

Inference in the Project Graph

Declarestack.jsoninference

RunInferenceintelligence

MeterOne credit balanceEU zone

stack.json — inference configuration plan · deploy

{
  "resources": {
    "inference_endpoints": [{ "name": "classifier", "runtime": "python",
      "model_bucket": "models", "model_prefix": "classifier-v3/" }],
    "workflows": [{ "name": "triage", "steps": [
      { "id": "score", "ai": { "model": "sentiment-multilingual-v1",
          "prompt": "Classify the tone of this message.", "output_schema": { "type": "object" } } }
    ] }]
  }
}

Metering · catalog models and compute 1 Shared Balance

text-embed-v1 is 3 credits per 1,000 tokens. sentiment-multilingual-v1 is 2. Vision catalog models are 3 credits per image. Compute is 0.01 credits per GiB-second.

Replaces: Amazon SageMaker · Workers AI · no direct equivalent

FAQ

Frequently asked questions

Which models are in the catalogue?

The named starting set: text-embed-v1 (embeddings, 3 credits/1k tokens), sentiment-multilingual-v1 (2 credits/1k tokens), and vision models (3 credits per image). Anything beyond that is a custom endpoint you declare.

Can we run our own fine-tuned model?

Yes — declare an inference endpoint with its runtime (and custom image if needed) and its model bucket/prefix. It meters on compute while it works and is invoked like any catalogue model.

Where does inference spend appear?

On the project’s single credit balance, attributed to the code that called it. “Which feature spent the model budget” is a reading of your own project, not a vendor-console excavation.

Connected in the Project Graph

Related capabilities, architectures, and comparisons

Ship your stack on the Software Lifecycle Cloud.

Closed preview on dedicated European infrastructure. Bring a repository or start from a blueprint.

Talk to us See pricing & credits →