Models

You select a model by the model string on InferenceInput. Mozaik resolves that name against the runtime's supportedModels list, maps ModelContext to the provider API, and returns typed ContextItems.

The default list is the exported supportedModels array. Each entry is a GenerativeModel: an Endpoint plus a ModelSpecification.

Bundled names

Credentials are read from the environment — see Installation. DeepSeek uses OPENAI_API_KEY / OPENAI_BASE_URL pointed at DeepSeek.

The exported endpoint classes (OpenAIResponses, OpenAIChatCompletions, AnthropicMessages, GeminiGenerateContent) are the adapters behind those rows. You only need them if you build a custom supportedModels list.

Specification flags

Each model has a ModelSpecification:

Requesting a feature the specification rejects fails validation before the API is called.

DeepSeek models set supportsStructuredOutput: false. The bundled OpenAI, Anthropic, and Gemini models set streaming, function calling, and structured output to true.

Custom runner

Pass inferenceRunnerConfig to initializeRuntime to replace the default list or the runner itself:

import { defineRuntime, RuntimeState, DefaultInferenceRunner, supportedModels } from '@mozaik-ai/core';

class AppState extends RuntimeState {}

const { initializeRuntime } = defineRuntime<AppState>();

initializeRuntime({
  state: new AppState(),
  inferenceRunnerConfig: {
    supportedModels, // or a subset / your own GenerativeModel[]
    // runner: new DefaultInferenceRunner(supportedModels, validator),
  },
});

InferenceRunner is { run(request): Promise<InferenceOutput>; stream(request): AsyncGenerator<SemanticEvent> }. DefaultInferenceRunner is the built-in implementation.

OpenResponses

ContextItem types follow the OpenResponses vocabulary: the same conceptual items for context and responses even when wire formats differ. Swapping providers is a model string change, not a domain-model change. Persistence stays typed items, not vendor-specific JSON.