The DeepSeek model driver for Namzu.

Install · Usage · Image input · Thinking mode · Documentation

Implements the kernel's LLMProvider over DeepSeek's Chat Completions endpoint, which is OpenAI-shaped — and then diverges where it counts, which is why this is its own package rather than a baseURL on @namzu/openai. Installed only if you use it; the kernel has no preferred vendor.

Install

pnpm add @namzu/sdk @namzu/deepseek

@namzu/sdk is a peer dependency. Install both. Requires Node.js 20+.

Usage

import { ProviderRegistry } from '@namzu/sdk'
import { registerDeepSeek } from '@namzu/deepseek'

registerDeepSeek() // once, at startup

const { provider } = ProviderRegistry.create({
  type: 'deepseek',
  apiKey: process.env.DEEPSEEK_API_KEY ?? '',
  model: 'deepseek-v4-flash',
})

for await (const chunk of provider.chatStream({
  model: 'deepseek-v4-flash',
  messages: [{ role: 'user', content: 'Is 91 prime?' }],
})) {
  if (chunk.delta.reasoning?.text) process.stderr.write(chunk.delta.reasoning.text)
  if (chunk.delta.content) process.stdout.write(chunk.delta.content)
}

The built-in catalogue offers deepseek-v4-flash, deepseek-v4-pro, and deepseek-v4-flash-vision-exp. deepseek-chat and deepseek-reasoner were discontinued on 2026-07-24 and are not aliases.

Image input

Only deepseek-v4-flash-vision-exp accepts images. User attachments are sent as ordered image_url data-URL parts. Images returned by tools keep their role: "tool" result and follow it in a separate user image message, so the tool-call/result sequence remains valid. PNG, JPEG, WebP and GIF are accepted.

The SDK's accumulated request-rich-content budget still applies before this projection and leaves durable Run.messages unchanged. This driver currently uses inline base64 only; it does not upload through the Files API or claim to normalize or decode caller-supplied image bytes. Documents, unresolved stored refs, unsupported image types, and images sent to either text model are refused before transport.

Thinking mode

Thinking is on by default — the vendor's default, not this driver's. The chain of thought arrives as delta.reasoning fragments, the same channel @namzu/anthropic uses. The kernel persists it with versioned adapter state and the exact provider/model/fallback route, so the same configured route replays it after restart or /resume. A model, provider, or chain-member switch keeps the portable assistant/tool history and omits the foreign native reasoning field.

Turn it off per call with thinking: { type: 'disabled' }.

Two parameters are refused rather than sent, because the vendor accepts them and applies neither: effort, and the sampling parameters while thinking is on. Both refusals, and how to opt out, are in the documentation below.

Documentation

• The DeepSeek driver — image input, thinking mode, reasoning replay, what this wire refuses, and why it carries no price rows
• Namzu docs

License

FSL-1.1-MIT, converting to MIT two years after each release.