DeepSeek Provider Guide

Text and image input with DeepSeek-V4.1-Flash and DeepSeek-V4-Pro through a single API

Overview

DeepSeek is a Chinese AI research lab offering highly capable open-weight models via a hosted cloud API. NeuroLink wraps their OpenAI-compatible endpoint. The API serves two models:

deepseek-flash — DeepSeek-V4.1-Flash. Reads images as well as text.
deepseek-v4-pro — DeepSeek-V4-Pro-0813. Text only.

The older ids still work as aliases of deepseek-flash: deepseek-chat answers with thinking off, and deepseek-reasoner answers with thinking on and returns a reasoning trace.

Key Facts

Protocol: OpenAI-compatible (/v1/chat/completions)
Default base URL: https://api.deepseek.com
Context window: 1M tokens (1,048,576), with up to 384K (393,216) output tokens
Vision: deepseek-flash and its aliases; not deepseek-v4-pro
Streaming: Supported
Tool calling: Supported. Tools and JSON output work in the same request.
Reasoning trace: returned as reasoning_content when thinking is on (deepseek-reasoner)

Quick Start

1. Get an API Key

2. Configure Environment

Add to your .env file:

# Required
DEEPSEEK_API_KEY=sk-...

# Optional: override the default model (default: deepseek-chat)
DEEPSEEK_MODEL=deepseek-chat

# Optional: override the base URL (default: https://api.deepseek.com)
DEEPSEEK_BASE_URL=https://api.deepseek.com

3. Install NeuroLink

npm install @juspay/neurolink
# or
pnpm add @juspay/neurolink

4. Generate Your First Response

import { NeuroLink } from "@juspay/neurolink";

const ai = new NeuroLink();

const result = await ai.generate({
  provider: "deepseek",
  input: {
    text: "Explain the difference between synchronous and asynchronous programming.",
  },
});

console.log(result.content);

Supported Models

Model ID	Serves	Context	Images	Notes
`deepseek-chat`	DeepSeek-V4.1-Flash	1M	Yes	Default; alias of `deepseek-flash`, thinking off
`deepseek-reasoner`	DeepSeek-V4.1-Flash	1M	Yes	Alias of `deepseek-flash`, thinking on
`deepseek-flash`	DeepSeek-V4.1-Flash	1M	Yes	Thinking on by default
`deepseek-v4-pro`	DeepSeek-V4-Pro	1M	No	Thinking on by default

Pass any model ID via --model (CLI) or model: (SDK). GET /models on api.deepseek.com lists only deepseek-flash and deepseek-v4-pro; the two older ids are accepted as aliases.

SDK Usage

Basic Generation

import { NeuroLink } from "@juspay/neurolink";

const ai = new NeuroLink();

const result = await ai.generate({
  provider: "deepseek",
  input: { text: "Write a TypeScript function to debounce an async function." },
});

console.log(result.content);

Using the Reasoner Model

const result = await ai.generate({
  provider: "deepseek",
  model: "deepseek-reasoner",
  input: { text: "Prove that the square root of 2 is irrational." },
});

// The reasoning trace is available separately from the final answer
console.log(result.content);

Note: with thinking on, responses take longer because the model reasons before answering.

Streaming

import { NeuroLink } from "@juspay/neurolink";

const ai = new NeuroLink();

const stream = await ai.stream({
  provider: "deepseek",
  input: { text: "Explain how B-trees work, step by step." },
});

for await (const chunk of stream.stream) {
  process.stdout.write(chunk);
}

Per-Call Credential Override

Pass credentials at call time to override the instance-level or environment-variable defaults. Useful when routing requests for different users through separate DeepSeek accounts.

const result = await ai.generate({
  provider: "deepseek",
  input: { text: "Hello, world!" },
  credentials: {
    deepseek: {
      apiKey: "sk-user-specific-key",
    },
  },
});

You can also override the base URL per call — useful when pointing at a self-hosted OpenAI-compatible proxy in front of DeepSeek:

const result = await ai.generate({
  provider: "deepseek",
  input: { text: "Hello" },
  credentials: {
    deepseek: {
      apiKey: "sk-...",
      baseURL: "https://my-proxy.example.com/v1",
    },
  },
});

CLI Usage

Basic Commands

# Generate with default model (deepseek-chat)
pnpm run cli generate "What is the halting problem?" --provider deepseek

# Use an alias
pnpm run cli generate "Hello" --provider ds

# Use the reasoning model
pnpm run cli generate "Prove P != NP (attempt)" --provider deepseek --model deepseek-reasoner

# Interactive loop mode
pnpm run cli loop --provider deepseek

Streaming via CLI

The CLI streams output by default when a TTY is attached. No extra flags are required.

pnpm run cli generate "Explain TCP/IP in detail" --provider deepseek --model deepseek-chat

Provider Aliases

The DeepSeek provider can be referenced by any of the following names:

Alias	Example
`deepseek`	`--provider deepseek`
`ds`	`--provider ds`

Configuration Reference

Environment Variable	Required	Default	Description
`DEEPSEEK_API_KEY`	Yes	—	DeepSeek API key (starts with `sk-`)
`DEEPSEEK_MODEL`	No	`deepseek-chat`	Default model to use
`DEEPSEEK_BASE_URL`	No	`https://api.deepseek.com`	Base URL for the API (override for proxies)

Feature Support Matrix

Feature	`deepseek-chat`	`deepseek-reasoner`	`deepseek-flash`	`deepseek-v4-pro`
Text generation	Yes	Yes	Yes	Yes
Streaming	Yes	Yes	Yes	Yes
Tool calling	Yes	See below	See below	See below
Structured output	Yes	Not tested	Not tested	Not tested
Vision / images	Yes	Yes	Yes	No
Embeddings	No	No	No	No
Reasoning trace	No	Yes	Yes	Yes

Troubleshooting

"Invalid DeepSeek API key"

The DEEPSEEK_API_KEY is missing or incorrect.

# Verify the variable is set
echo $DEEPSEEK_API_KEY

# Set it inline
export DEEPSEEK_API_KEY=sk-...

Get or rotate keys at https://platform.deepseek.com/api_keys.

"DeepSeek account has insufficient balance"

Your account credit is exhausted. Top up at https://platform.deepseek.com/usage.

"DeepSeek rate limit exceeded"

Too many requests in a short window. Implement exponential backoff or reduce request concurrency. Rate limits are published in the DeepSeek API docs.

"Model not found"

DeepSeek answers an unknown id with 400 and "The supported API model names are deepseek-flash, deepseek-v4-pro". Use one of those, or the deepseek-chat / deepseek-reasoner aliases.

Slow responses with thinking on

Expected. deepseek-reasoner, deepseek-flash and deepseek-v4-pro reason before answering, which adds latency. Use deepseek-chat for latency-sensitive paths.

Tool calls failing with thinking on

DeepSeek's thinking mode guide says each turn's reasoning_content must be sent back on every later request once tools are in play, and that the API returns 400 otherwise. NeuroLink does this for you: the DeepSeek catalog entry sets quirks.replayReasoningContent, so both generate() and stream() send the reasoning back on each follow-up request. (Tested live on 2026-09-26, the API accepted follow-ups with or without it, but NeuroLink follows the documented contract.)

If a tool loop still fails, check whether you are replaying conversation history yourself without each assistant turn's reasoning. For tool-heavy workflows where you don't need the reasoning, deepseek-chat keeps thinking off.

Overview​

Key Facts​

Quick Start​

1. Get an API Key​

2. Configure Environment​

3. Install NeuroLink​

4. Generate Your First Response​

Supported Models​

SDK Usage​

Basic Generation​

Using the Reasoner Model​

Streaming​

Per-Call Credential Override​

CLI Usage​

Basic Commands​

Streaming via CLI​

Provider Aliases​

Configuration Reference​

Feature Support Matrix​

Troubleshooting​

"Invalid DeepSeek API key"​

"DeepSeek account has insufficient balance"​

"DeepSeek rate limit exceeded"​

"Model not found"​

Slow responses with thinking on​

Tool calls failing with thinking on​

See Also​