DeepSeek just released the [V4 model family](https://api-docs.deepseek.com/news/news260424), featuring V4 Pro and V4 Flash. Both support a 1M context window and pack significantly stronger Agent capabilities.
Let's get them running in Yao Agents.
## Find Your App Directory
All config files live in the Yao Engine app directory. Open Yao Engine and click **Open Folder** to get there.
You'll see a structure like this:
```
├── connectors/ ← Connector config files go here
│ ├── default.conn.yao
│ ├── thinking.conn.yao
│ ├── vision.conn.yao
│ └── deepseek/ ← Create this subdirectory
├── assistants/
├── .env
└── ...
```
## Add the DeepSeek V4 Connectors
Create a `deepseek/` directory under `connectors/`, then add these four files.
### V4 Pro (Non-thinking Mode)
File: `connectors/deepseek/v4-pro.conn.yao`
```json
{
"label": "DeepSeek V4 Pro",
"type": "openai",
"options": {
"model": "deepseek-v4-pro",
"key": "$ENV.DEEPSEEK_V4_API_KEY",
"proxy": "$ENV.DEEPSEEK_V4_PROXY",
"thinking": {
"type": "disabled"
},
"capabilities": {
"tool_calls": true,
"streaming": true,
"json": true
}
}
}
```
### V4 Pro Thinking (Thinking Mode)
File: `connectors/deepseek/v4-pro-thinking.conn.yao`
```json
{
"label": "DeepSeek V4 Pro Thinking",
"type": "openai",
"options": {
"model": "deepseek-v4-pro",
"key": "$ENV.DEEPSEEK_V4_API_KEY",
"proxy": "$ENV.DEEPSEEK_V4_PROXY",
"thinking": {
"type": "enabled"
},
"capabilities": {
"tool_calls": true,
"reasoning": true,
"streaming": true,
"json": true
}
}
}
```
### V4 Flash (Non-thinking Mode)
File: `connectors/deepseek/v4-flash.conn.yao`
```json
{
"label": "DeepSeek V4 Flash",
"type": "openai",
"options": {
"model": "deepseek-v4-flash",
"key": "$ENV.DEEPSEEK_V4_API_KEY",
"proxy": "$ENV.DEEPSEEK_V4_PROXY",
"thinking": {
"type": "disabled"
},
"capabilities": {
"tool_calls": true,
"streaming": true,
"json": true
}
}
}
```
### V4 Flash Thinking (Thinking Mode)
File: `connectors/deepseek/v4-flash-thinking.conn.yao`
```json
{
"label": "DeepSeek V4 Flash Thinking",
"type": "openai",
"options": {
"model": "deepseek-v4-flash",
"key": "$ENV.DEEPSEEK_V4_API_KEY",
"proxy": "$ENV.DEEPSEEK_V4_PROXY",
"thinking": {
"type": "enabled"
},
"capabilities": {
"tool_calls": true,
"reasoning": true,
"streaming": true,
"json": true
}
}
}
```
### Set Up Environment Variables
Add the following to the `.env` file in your app directory:
```bash
DEEPSEEK_V4_API_KEY=sk-your-deepseek-api-key
DEEPSEEK_V4_PROXY=https://api.deepseek.com
```
Grab your API key from the [DeepSeek platform](https://platform.deepseek.com/).
Once you've added the files, restart Yao Engine. The four DeepSeek V4 connectors will show up in the model picker for both AI Expert and Robot.
## Using DeepSeek V4 in the Sandbox
Yao Agents Sandbox supports two Runners — `claude` and `opencode`. DeepSeek V4 works with both. The `sandbox.yao` file lives in each expert's configuration directory at `assistants/<namespace>/<assistant-name>/sandbox.yao`.
### Specifying a Vision Model
DeepSeek V4 doesn't support vision (image understanding) yet, but some Sandbox tasks need to analyze screenshots. The `opencode` Runner lets you set a separate vision connector in `sandbox.yao`:
```json
{
"version": "2.0",
"computer": {
"image": "yaoapp/tai-sandbox-opencode:latest",
"memory": "2GB",
"cpus": 2,
"work_dir": "/workspace"
},
"runner": {
"name": "opencode",
"mode": "cli",
"options": {
"permission_mode": "bypassPermissions"
},
"connectors": {
"vision": "openai.gpt-4o-mini"
}
},
"lifecycle": "oneshot"
}
```
Key settings:
- `runner.name` — set to `"opencode"`
- `runner.connectors.vision` — point this to a vision-capable model connector (e.g. `openai.gpt-4o-mini`). Screenshots go through this model; everything else uses DeepSeek V4.
### opencode Runner Docker Images
The `opencode` Runner ships with several Docker images. Pick the one that fits your use case:
| Image | Desktop | SSH | Includes Yao | Best For |
|------|---------|-----|----------|---------|
| `tai-sandbox-opencode` | — | — | — | Pure code tasks, most use cases |
| `tai-sandbox-opencode-ssh` | — | ✓ | — | Remote server access |
| `tai-sandbox-opencode-desktop-lite` | Lightweight | — | — | Tasks that need a browser |
| `tai-sandbox-opencode-desktop-lite-ssh` | Lightweight | ✓ | — | Browser + remote server |
| `tai-sandbox-opencode-desktop` | Full (XFCE) | — | — | Applet-style tasks |
| `tai-sandbox-opencode-desktop-ssh` | Full (XFCE) | ✓ | — | Site building, deployment |
| `tai-sandbox-opencode-yao` | — | ✓ | ✓ | Yao app development |
| `tai-sandbox-opencode-yao-desktop` | Full (XFCE) | ✓ | ✓ | Yao app dev (with desktop) |
| `tai-sandbox-opencode-yao-desktop-lite` | Lightweight | ✓ | ✓ | Yao app dev (lightweight desktop) |
Pull an image:
```bash
docker pull yaoapp/tai-sandbox-opencode:latest
```
Swap in a different image name as needed. All images support both `linux/amd64` and `linux/arm64`.
## V4 Pro vs. V4 Flash
| | V4 Pro | V4 Flash |
|---|--------|----------|
| Reasoning | Top-tier, on par with closed-source models | Close to Pro |
| Agent capability | Best in open-source, excels at complex tasks | Comparable to Pro on simple tasks, falls behind on complex ones |
| World knowledge | Rich, second only to Gemini-Pro-3.1 | Somewhat weaker |
| Speed & cost | Slower, pricier | Faster, cheaper |
| Context window | 1M | 1M |
Rule of thumb: use V4 Pro Thinking for complex Agent tasks, V4 Flash for everyday work.
## Further Reading
- [AI Model Configuration Guide](/docs/en-us/settings/ai-models) — Connector file format, environment variables, and more examples
- [DeepSeek V4 Official Release Notes](https://api-docs.deepseek.com/news/news260424) — Full technical details and API docs
tutorial
Setting Up DeepSeek V4 in Yao Agents
DeepSeek V4 just dropped — 1M context window, massively improved Agent capabilities. Here's how to hook up DeepSeek V4 Pro and V4 Flash in Yao Agents.