Run OpenClaw Agents with DeepSeek V3
DeepSeek V3.2 was purpose-built for agentic workflows — trained with a specialized pipeline to handle multi-step tool use and complex reasoning chains. This guide shows you how to run OpenClaw agents with DeepSeek V3, both locally via Ollama and through DeepSeek's cheap API.
Why DeepSeek V3 for OpenClaw
Most LLMs are trained primarily on conversation and instruction following. DeepSeek V3.2 used a large-scale agentic task synthesis pipeline during training — it was specifically optimized for multi-step tool use, where decisions at step 3 depend on results from step 1 and 2. This is exactly the kind of reasoning OpenClaw agents need.
The result: DeepSeek V3.2 maintains coherent plans across long tool-calling chains, handles unexpected tool results without losing its thread, and produces more reliable outputs on complex agentic tasks than models of similar capability trained without this focus.
DeepSeek V3 Variants: Which One to Use
| Model | Size | How to run | Best for |
|---|---|---|---|
| DeepSeek V3 API | 685B (full) | api.deepseek.com | Best reasoning, cheap API |
| deepseek-v3.2 (Ollama) | 685B Q4 | ollama pull deepseek-v3.2 | Server hardware only |
| deepseek-v3.1 (Ollama) | 685B Q4 | ollama pull deepseek-v3.1 | Server hardware, stable release |
| deepseek-v3 (Ollama) | 685B Q2 | ollama pull deepseek-v3 | RTX 4090 feasible (slower) |
Practical recommendation: For most users, DeepSeek V3 API is the right path. The full model is far too large for consumer hardware. The API gives you the full 685B model at $0.27/M tokens — significantly cheaper than GPT-4o or Claude Sonnet. Use local Ollama only if you have server hardware or strict data privacy requirements.
Option A: DeepSeek via API (Recommended)
DeepSeek offers cheap API access to their full V3.2 model. This is the most practical path for most OpenClaw users.
Step 1: Get a DeepSeek API key
# 1. Sign up at: platform.deepseek.com
# 2. Go to API Keys → Create new key
# 3. Set the key in your environment
export DEEPSEEK_API_KEY="sk-..."
# DeepSeek API is OpenAI-compatible
# Base URL: https://api.deepseek.comStep 2: Configure OpenClaw for DeepSeek API
{
"models": {
"providers": {
"deepseek": {
"baseUrl": "https://api.deepseek.com/v1",
"apiKey": "${DEEPSEEK_API_KEY}",
"api": "openai",
"models": [
{
"id": "deepseek-chat",
"name": "DeepSeek V3",
"contextWindow": 64000,
"maxOutput": 8192,
"toolCalling": true
},
{
"id": "deepseek-reasoner",
"name": "DeepSeek R1",
"contextWindow": 64000,
"maxOutput": 8192,
"toolCalling": true
}
]
}
}
}
}Step 3: Set model in SOUL.md
## Identity
- Name: Orion
- Role: Project Manager
- Model: deepseek/deepseek-chat ← DeepSeek V3
- Timezone: UTC
# For reasoning-heavy tasks, use R1:
- Model: deepseek/deepseek-reasonerOption B: DeepSeek via Ollama (Local)
If you have a high-end GPU (RTX 4090 or similar with 24GB VRAM) or server hardware, you can run quantized DeepSeek V3 locally.
# Pull DeepSeek V3.2 (requires significant storage)
ollama pull deepseek-v3.2
# Or the previous stable release
ollama pull deepseek-v3.1
# On RTX 4090 with Q4 quantization: ~18-25 tokens/sec
# On Apple Silicon M4 Max (128GB): feasible but slow
# Configure openclaw.json for local DeepSeek
{
"providers": {
"ollama": {
"baseUrl": "http://127.0.0.1:11434",
"api": "ollama",
"models": [{
"id": "deepseek-v3.2",
"contextWindow": 64000,
"toolCalling": true
}]
}
}
}
# SOUL.md model:
- Model: ollama/deepseek-v3.2Data privacy note: If you are running agents on sensitive business data, local Ollama keeps everything on your machine. DeepSeek API sends data to DeepSeek's servers. For confidential or legally sensitive workflows, local is the safer path.
API Cost: DeepSeek vs Competitors
| Model | Input / 1M tokens | Output / 1M tokens | vs DeepSeek V3 |
|---|---|---|---|
| DeepSeek V3 | $0.27 | $1.10 | baseline |
| Gemini 2.0 Flash | $0.10 | $0.40 | Cheaper, less reasoning |
| Claude Haiku 4.5 | $0.80 | $4.00 | 3x pricier |
| GPT-4o Mini | $0.15 | $0.60 | Similar price, less power |
| Claude Sonnet 4.5 | $3.00 | $15.00 | 11x pricier |
For reasoning-heavy agents that need complex multi-step tool use, DeepSeek V3 API delivers near-frontier reasoning at a fraction of Claude Sonnet's cost. This makes it compelling for 24/7 agents where API costs accumulate.
Best Agent Types for DeepSeek V3
Complex multi-step research agents
Best fitDeepSeek V3.2 was specifically trained for multi-step agentic workflows. Agents that need to plan, search, synthesize, and iterate perform particularly well.
Project management and decision agents
Best fitAgents that weigh tradeoffs, prioritize tasks, and make decisions based on multiple inputs benefit from DeepSeek V3's strong reasoning. Good for Orion-style PM agents.
Code analysis and review agents
Strong fitStrong on technical reasoning, code understanding, and structured output. Pair with the files skill for codebase-level analysis.
Budget-conscious production agents
Strong fitFor agents running 24/7 with high API call volume, DeepSeek V3's pricing makes it 3-10x cheaper than Claude alternatives at similar reasoning quality.
Known Limitations
⚠ Full model requires server hardware locally
685B parameters cannot run on a MacBook or gaming PC. For local use, you need 140GB+ RAM or multiple GPUs. The API is the practical path for most users.
⚠ Data privacy consideration
DeepSeek is a Chinese company. API usage means your conversation data is processed on their infrastructure. For sensitive data, use local inference or a different provider.
⚠ Context window smaller than Gemma 4
DeepSeek V3's effective context window is 64K tokens via the API, compared to 256K for Gemma 4 26B. For very long document processing, Gemma 4 has an advantage.
⚠ No free tier
DeepSeek's API is paid (though cheap). Unlike Google AI Studio's free Gemma/Gemini tier, there is no zero-cost option for the API. Budget for API costs accordingly.
Related Guides
Frequently Asked Questions
Can DeepSeek V3 run locally on consumer hardware?
The full DeepSeek V3.2 model (685B parameters) requires server-grade hardware — 140GB+ RAM or multiple high-end GPUs. For consumer hardware, use the DeepSeek V3.1 or V3 smaller distilled variants available on Ollama, or use DeepSeek's API which provides cheap access to the full model. On an RTX 4090 (24GB VRAM), the Q4 quantized V3 runs at 18-25 tokens/second.
What makes DeepSeek V3.2 good for agent tasks?
DeepSeek V3.2 was trained with a large-scale agentic task synthesis pipeline — it was specifically optimized for multi-step tool use, not just general language tasks. This shows in practice: it maintains coherent reasoning across long tool-calling chains and handles situations where intermediate tool results affect subsequent decisions.
How much does the DeepSeek API cost compared to Claude?
DeepSeek V3 API costs approximately $0.27 per million input tokens and $1.10 per million output tokens. Claude Haiku 4.5 costs $0.80/$4.00. Claude Sonnet costs $3/$15. For high-volume agent workloads, DeepSeek's API is 3-10x cheaper than comparable Anthropic models while delivering strong reasoning performance.
Is DeepSeek safe to use for business data?
DeepSeek is a Chinese company and their cloud API means conversation data is processed on their servers. For sensitive business data, legal documents, or anything subject to data residency requirements, run DeepSeek locally via Ollama instead. Local inference keeps all data on your machine. For non-sensitive workloads, the API is fine.
How does DeepSeek V3 compare to Qwen3 for OpenClaw agents?
Qwen3 30B-A3B is better for consumer hardware (20GB RAM vs 140GB+ for full V3.2). DeepSeek V3.2's full model outperforms Qwen3 on complex multi-step reasoning tasks when run via API. For local deployment, Qwen3 is the more practical choice. For API-based agents where you want strong reasoning at low cost, DeepSeek V3 API is compelling.
Agent templates optimized for reasoning-heavy workflows
CrewClaw templates include pre-tuned Rules sections for complex multi-step agent tasks. Download, set your DeepSeek API key, and deploy in minutes.
Deploy a Ready-Made AI Agent
Skip the setup. Pick a template and deploy in 60 seconds.