AI models

Understand how language and image models are configured, and how to change the default chat model.

All model configuration lives in lib/ai/models.ts. The app uses a custom provider built on top of the Vercel AI SDK to group models under stable internal identifiers, decoupling the rest of the codebase from specific upstream model names.

Model registry

The myProvider object registers every model the app can use:

lib/ai/models.ts

import { openai } from '@ai-sdk/openai';
import { fireworks } from '@ai-sdk/fireworks';
import {
  customProvider,
  extractReasoningMiddleware,
  wrapLanguageModel,
} from 'ai';

export const DEFAULT_CHAT_MODEL: string = 'chat-model-small';

export const myProvider = customProvider({
  languageModels: {
    'chat-model-small': openai('gpt-4o-mini'),
    'chat-model-large': openai('gpt-4o'),
    'chat-model-reasoning': wrapLanguageModel({
      model: fireworks('accounts/fireworks/models/deepseek-r1'),
      middleware: extractReasoningMiddleware({ tagName: 'think' }),
    }),
    'title-model': openai('gpt-4-turbo'),
    'artifact-model': openai('gpt-4o-mini'),
  },
  imageModels: {
    'small-model': openai.image('dall-e-2'),
    'large-model': openai.image('dall-e-3'),
  },
});

Language models

Internal ID	Upstream model	Purpose
chat-model-small	gpt-4o-mini	Default model for everyday conversations. Fast and cost-efficient.
chat-model-large	gpt-4o	Full-size model for complex, multi-step tasks.
chat-model-reasoning	deepseek-r1 (via Fireworks)	Step-by-step reasoning for problems that benefit from chain-of-thought.
title-model	gpt-4-turbo	Auto-generates a short title for each new chat. Not user-selectable.
artifact-model	gpt-4o-mini	Generates documents and code artifacts in the artifact panel. Not user-selectable.

The three models exposed in the chat UI (chat-model-small, chat-model-large, chat-model-reasoning) are defined in the chatModels array, which drives the model selector component:

lib/ai/models.ts

export const chatModels: Array<ChatModel> = [
  {
    id: 'chat-model-small',
    name: 'Small model',
    description: 'Small model for fast, lightweight tasks',
  },
  {
    id: 'chat-model-large',
    name: 'Large model',
    description: 'Large model for complex, multi-step tasks',
  },
  {
    id: 'chat-model-reasoning',
    name: 'Reasoning model',
    description: 'Uses advanced reasoning',
  },
];

Image models

Internal ID	Upstream model	Characteristics
small-model	dall-e-2	Faster generation, lower cost.
large-model	dall-e-3	Higher quality, supports more detailed prompts.

Reasoning middleware

The chat-model-reasoning model wraps DeepSeek R1 with extractReasoningMiddleware. This middleware strips the model’s internal <think>...</think> chain-of-thought from the streamed output before it reaches the UI, while still making the reasoning trace available separately.

'chat-model-reasoning': wrapLanguageModel({
  model: fireworks('accounts/fireworks/models/deepseek-r1'),
  middleware: extractReasoningMiddleware({ tagName: 'think' }),
}),

Using chat-model-reasoning requires a valid FIREWORKS_API_KEY in your environment. See Environment variables.

Changing the default model

The DEFAULT_CHAT_MODEL constant controls which model is pre-selected when a user starts a new chat:

lib/ai/models.ts

export const DEFAULT_CHAT_MODEL: string = 'chat-model-small';

To change the default, replace 'chat-model-small' with any other user-facing model ID:

export const DEFAULT_CHAT_MODEL: string = 'chat-model-large';

Valid values are 'chat-model-small', 'chat-model-large', and 'chat-model-reasoning'.

Swapping to a different upstream model

To point an internal ID at a different upstream model, edit the corresponding entry in myProvider. For example, to upgrade chat-model-small to a newer OpenAI model:

'chat-model-small': openai('gpt-4o'),

The rest of the codebase references only the internal ID, so no other files need to change.

Swapping models may affect response quality, latency, and API costs. Test thoroughly before deploying to production.