DeepSeek

The DeepSeek provider offers access to powerful language models through the DeepSeek API.

API keys can be obtained from the DeepSeek Platform.

Setup

The DeepSeek provider is available via the @ai-sdk/deepseek module. You can install it with:

$ pnpm add @ai-sdk/deepseek

$ npm install @ai-sdk/deepseek

$ yarn add @ai-sdk/deepseek

$ bun add @ai-sdk/deepseek

Provider Instance

You can import the default provider instance deepSeek from @ai-sdk/deepseek:

import { deepSeek } from '@ai-sdk/deepseek';

For custom configuration, you can import createDeepSeek and create a provider instance with your settings:

import { createDeepSeek } from '@ai-sdk/deepseek';

const deepSeek = createDeepSeek({
  apiKey: process.env.DEEPSEEK_API_KEY ?? '',
});

You can use the following optional settings to customize the DeepSeek provider instance:

• baseURL string

  Use a different URL prefix for API calls. The default prefix is https://api.deepseek.com.

• apiKey string

  API key that is being sent using the Authorization header. It defaults to the DEEPSEEK_API_KEY environment variable.

• headers Record<string,string>

  Custom headers to include in the requests.

• fetch (input: RequestInfo, init?: RequestInit) => Promise<Response>

  Custom fetch implementation.

Language Models

You can create language models using a provider instance:

import { deepSeek } from '@ai-sdk/deepseek';
import { generateText } from 'ai';

const { text } = await generateText({
  model: deepSeek('deepseek-v4-flash'),
  prompt: 'Write a vegetarian lasagna recipe for 4 people.',
});

You can also use the .chat() or .languageModel() factory methods:

const model = deepSeek.chat('deepseek-v4-flash');
// or
const model = deepSeek.languageModel('deepseek-v4-flash');

DeepSeek language models can be used in the streamText function (see AI SDK Core).

DeepSeek retired the deepseek-chat and deepseek-reasoner aliases on July 24, 2026. Use deepseek-flash (the alias for the current V4.x Flash release), deepseek-v4-flash, or deepseek-v4-pro for the current API. Custom and legacy model IDs remain accepted as strings for compatibility with custom endpoints.

The following optional provider options are available for DeepSeek models:

• logprobs boolean

  Optional. Returns log probabilities for generated content and reasoning tokens in providerMetadata.deepseek.logprobs.

• topLogprobs number

  Optional. Returns the specified number of most likely tokens at each token position. Accepts values from 0 through 20 and automatically enables logprobs.

• userId string

  Optional. An opaque end-user identifier that DeepSeek uses for content-safety tracing, KV-cache isolation, and scheduling isolation. The value must match ^[a-zA-Z0-9_-]+$ and contain at most 512 characters. Do not include names, email addresses, or other private user information.

• thinking object

  Optional. Controls thinking mode (chain-of-thought reasoning) for DeepSeek V4 models.

  • type: 'enabled' | 'disabled' - Enable or disable thinking mode. See DeepSeek's thinking mode docs.

• reasoningEffort 'low' | 'high' | 'max'

  Optional. Controls thinking strength for DeepSeek V4 reasoning models. When using the top-level reasoning setting, minimal is sent as low, medium is sent as high, xhigh is sent as max, and max is sent directly. A compatibility warning is returned whenever the requested value is mapped.

For backwards compatibility, legacy provider options supplied at runtime are also mapped to documented values: thinking.type: 'adaptive' becomes 'enabled', reasoningEffort: 'medium' becomes 'high', and reasoningEffort: 'xhigh' becomes 'max'. Each mapping returns a compatibility warning so callers can migrate to a canonical value.

DeepSeek has deprecated the top-level frequencyPenalty and presencePenalty settings. The provider omits these settings and returns a deprecation warning when they are used. temperature and topP have no effect while thinking is enabled, including the default thinking mode for DeepSeek V4 models, so the provider omits them with an unsupported warning. Explicitly set thinking.type to 'disabled' to use temperature and topP.

import {
  deepSeek,
  type DeepSeekLanguageModelChatOptions,
} from '@ai-sdk/deepseek';
import { generateText } from 'ai';

const { text, reasoning } = await generateText({
  model: deepSeek('deepseek-v4-flash'),
  prompt: 'How many "r"s are in the word "strawberry"?',
  providerOptions: {
    deepseek: {
      userId: 'tenant_123-user',
      thinking: { type: 'enabled' },
      reasoningEffort: 'high',
    } satisfies DeepSeekLanguageModelChatOptions,
  },
});

Message Names

DeepSeek supports optional participant names on system, user, and assistant messages. Set providerOptions.deepseek.name on each message that should include a name:

import {
  deepSeek,
  type DeepSeekMessageProviderOptions,
} from '@ai-sdk/deepseek';
import { generateText } from 'ai';

const { text } = await generateText({
  model: deepSeek('deepseek-flash'),
  instructions: {
    role: 'system',
    content: 'Help the customer plan a short trip.',
    providerOptions: {
      deepseek: {
        name: 'travel_planner',
      } satisfies DeepSeekMessageProviderOptions,
    },
  },
  messages: [
    {
      role: 'user',
      content: 'I want to visit Lisbon for a weekend.',
      providerOptions: {
        deepseek: {
          name: 'customer',
        } satisfies DeepSeekMessageProviderOptions,
      },
    },
    {
      role: 'assistant',
      content: 'What kinds of activities do you enjoy?',
      providerOptions: {
        deepseek: {
          name: 'travel_planner',
        } satisfies DeepSeekMessageProviderOptions,
      },
    },
    {
      role: 'user',
      content: 'Food, architecture, and walking.',
      providerOptions: {
        deepseek: {
          name: 'customer',
        } satisfies DeepSeekMessageProviderOptions,
      },
    },
  ],
});

The same message option works with streamText:

import {
  deepSeek,
  type DeepSeekMessageProviderOptions,
} from '@ai-sdk/deepseek';
import { streamText } from 'ai';

const result = streamText({
  model: deepSeek('deepseek-flash'),
  messages: [
    {
      role: 'user',
      content: 'Suggest a name for my neighborhood book club.',
      providerOptions: {
        deepseek: {
          name: 'organizer',
        } satisfies DeepSeekMessageProviderOptions,
      },
    },
  ],
});

for await (const textPart of result.textStream) {
  process.stdout.write(textPart);
}

Names are omitted when the option is not set. The name value must be a string. DeepSeek does not support names on tool messages, so the provider ignores that placement and returns an unsupported-feature warning. Avoid including unnecessary personal or identifying information in message names.

Reasoning

DeepSeek V4 models support reasoning. The reasoning is exposed through streaming:

import { deepSeek } from '@ai-sdk/deepseek';
import { streamText } from 'ai';

const result = streamText({
  model: deepSeek('deepseek-v4-pro'),
  prompt: 'How many "r"s are in the word "strawberry"?',
});

for await (const part of result.stream) {
  if (part.type === 'reasoning') {
    // This is the reasoning text
    console.log('Reasoning:', part.text);
  } else if (part.type === 'text') {
    // This is the final answer
    console.log('Answer:', part.text);
  }
}

See AI SDK UI: Chatbot for more details on how to integrate reasoning into your chatbot.

Chat Prefix Completion

DeepSeek's beta chat prefix completion continues the content of a final assistant message. Create a provider with a beta base URL and set prefix: true in that assistant message's provider options:

import {
  createDeepSeek,
  type DeepSeekAssistantMessageProviderOptions,
} from '@ai-sdk/deepseek';
import { generateText } from 'ai';

const deepSeek = createDeepSeek({
  baseURL: 'https://api.deepseek.com/beta',
});

const { text } = await generateText({
  model: deepSeek('deepseek-v4-flash'),
  messages: [
    {
      role: 'user',
      content: 'Write a short sentence about the color of the sky.',
    },
    {
      role: 'assistant',
      content: 'The sky is',
      providerOptions: {
        deepseek: {
          prefix: true,
        } satisfies DeepSeekAssistantMessageProviderOptions,
      },
    },
  ],
});

The prefixed message must be an assistant message and the final message in the prompt. The configured baseURL must end in /beta, including when using a proxy. Invalid placement or a non-beta base URL causes the request to fail before it is sent.

Strict Tool Calls

DeepSeek's strict tool-call mode is a beta feature. Create the provider with a beta base URL and set strict: true on every function tool in the request:

import { createDeepSeek } from '@ai-sdk/deepseek';
import { generateText, tool } from 'ai';
import { z } from 'zod';

const deepSeek = createDeepSeek({
  baseURL: 'https://api.deepseek.com/beta',
});

const result = await generateText({
  model: deepSeek('deepseek-flash'),
  prompt: 'What is the weather in San Francisco?',
  tools: {
    weather: tool({
      description: 'Get the weather for a location.',
      inputSchema: z.object({ location: z.string() }),
      strict: true,
      execute: async ({ location }) => ({ location, temperature: 18 }),
    }),
  },
});

Strict tools fail locally when the base URL does not end in /beta. When any function tool is strict, every function tool in the same request must set strict: true.

Provider Metadata

DeepSeek exposes the response system fingerprint and context cache usage through the providerMetadata property:

import { deepSeek } from '@ai-sdk/deepseek';
import { generateText } from 'ai';

const result = await generateText({
  model: deepSeek('deepseek-v4-flash'),
  prompt: 'Your prompt here',
});

console.log(result.providerMetadata);
// Example output:
// {
//   deepseek: {
//     systemFingerprint: 'fp_eaab8d114b_prod0820_fp8_kvcache',
//     promptCacheHitTokens: 1856,
//     promptCacheMissTokens: 5,
//   },
// }

The metadata includes:

• systemFingerprint: The backend configuration fingerprint for the response
• promptCacheHitTokens: Number of input tokens that were cached
• promptCacheMissTokens: Number of input tokens that were not cached

For streamed responses, the latest non-null fingerprint from the response chunks is returned.

Chat Response Metadata

DeepSeek preserves provider-specific response fields in providerMetadata.deepseek for generated and streamed responses:

• responseObject: chat.completion or chat.completion.chunk
• choiceIndex: the selected response choice index
• messageRole: the response message role, when supplied
• toolCallTypes: the tool-call type for each returned call

These fields remain in provider metadata because they are specific to DeepSeek's Chat Completions response rather than shared AI SDK result fields.

File Uploads

For inline images or image URLs, use file-part provider options to select the image processing detail. DeepSeek supports low, high, original, and auto:

import {
  deepSeek,
  type DeepSeekFilePartProviderOptions,
} from '@ai-sdk/deepseek';
import { generateText } from 'ai';

const { text } = await generateText({
  model: deepSeek('deepseek-v4-flash-vision-exp'),
  messages: [
    {
      role: 'user',
      content: [
        { type: 'text', text: 'Describe this image.' },
        {
          type: 'file',
          data: new URL('https://example.com/image.webp'),
          mediaType: 'image/webp',
          providerOptions: {
            deepseek: {
              imageDetail: 'low',
            } satisfies DeepSeekFilePartProviderOptions,
          },
        },
      ],
    },
  ],
});

Set fileData: true on an inline image file part to use DeepSeek's file_data content-part representation. This preserves the file part's filename. fileData cannot be used with image URLs or imageDetail.

DeepSeek accepts JPEG, PNG, GIF, and WebP image inputs. HTTP image URLs can be at most 8,192 characters.

You can upload images using the DeepSeek Files API and pass the returned provider reference to deepseek-v4-flash-vision-exp. This avoids sending the image bytes again in each request.

import { deepSeek, type DeepSeekFilesOptions } from '@ai-sdk/deepseek';
import { generateText, uploadFile } from 'ai';
import { readFile } from 'node:fs/promises';

const { providerReference, mediaType } = await uploadFile({
  api: deepSeek.files(),
  data: await readFile('./image.png'),
  filename: 'image.png',
  providerOptions: {
    deepseek: {
      expiresAfter: 3600,
    } satisfies DeepSeekFilesOptions,
  },
});

const { text } = await generateText({
  model: deepSeek('deepseek-v4-flash-vision-exp'),
  messages: [
    {
      role: 'user',
      content: [
        { type: 'text', text: 'Describe this image.' },
        {
          type: 'file',
          mediaType: mediaType ?? 'image/png',
          data: providerReference,
        },
      ],
    },
  ],
});

The optional expiresAfter setting specifies the lifetime in seconds. DeepSeek accepts values from 3,600 seconds (one hour) through 2,592,000 seconds (30 days). Files are permanent when no expiration is specified.

The provider requires a valid file ID in successful upload responses. It also validates returned object and purpose discriminators and numeric metadata. Other response metadata remains optional for compatibility with incomplete responses; omitted values are not included in providerMetadata, and an omitted response filename falls back to the filename supplied to uploadFile.

DeepSeek file uploads support JPEG (.jpg and .jpeg), PNG, GIF, and WebP images. Each file can be at most 64 MiB, and filenames can contain at most 512 characters. The AI SDK validates these constraints before sending the upload request. The image/jpg media type alias is accepted, and a supported filename extension is used as a fallback when the media type is generic (for example, application/octet-stream). Recognizable non-image content is rejected even when its declared media type or filename indicates a supported image format.

Model Capabilities