How to Use Qwen 3.5 in OpenClaw for Cost-Efficient AI Agents

OpenClaw has exploded in popularity as the go-to AI agent framework in 2026, but its default settings can burn through tokens fast. This guide shows how to deploy Qwen 3.5 Series (27B, 35B-A3B, 122B-A10B, 397B-A17B) via Novita AI to build powerful, cost-efficient agents that scale without breaking the bank.

What Is OpenClaw — and Why Is Everyone Using It?

OpenClaw (formerly Clawdbot) has become the most talked-about AI agent framework in 2026, dominating discussions across developer communities, enterprise AI teams, and tech media. Bloomberg called it “the AI marvel powering China’s agent boom,” while TechNode reported that Chinese firms are racing to adopt OpenClaw for everything from customer service bots to internal automation workflows.

• Multi-model routing: Connect to any LLM provider (OpenAI, Anthropic, local models) and dynamically route tasks to the best model for the job.
• Messaging platform integration: Deploy agents directly into Telegram, WhatsApp, Discord, iMessage, and more — no custom UI needed.
• Extensible skills system: Add capabilities like file operations, web search, code execution, and workflow automation through a plugin architecture.

But there’s a catch: OpenClaw is token-hungry. Default configurations can consume 5-10x more tokens than necessary, with features like heartbeat polling alone costing a lot on high-end models.

For developers running 24/7 agents or deploying multiple instances, token costs can spiral fast. The solution? Pair OpenClaw with cost-efficient, high-performance models like Qwen 3.5 Series on Novita AI.

What Is Qwen 3.5 Series — and Why Use It in OpenClaw?

Qwen 3.5 Series is Alibaba Cloud’s latest family of large language models, designed for production workloads that demand both intelligence and efficiency. The lineup includes four variants optimized for different use cases:

• Qwen 3.5 27B: Balanced reasoning and speed for general-purpose agent tasks.
• Qwen 3.5 35B-A3B: Enhanced parameter efficiency for complex multi-step reasoning.
• Qwen 3.5 122B-A10B: High-end capability for advanced reasoning, code generation, and long-context understanding.
• Qwen 3.5 397B-A17B: Maximum intelligence for the most demanding tasks requiring deep reasoning and comprehensive analysis.

Why Qwen 3.5 Series is Ideal for OpenClaw

When building AI agents with OpenClaw, you need models that deliver strong performance across reasoning, tool use, and instruction following — without excessive cost. Qwen 3.5 Series checks all the boxes:

1. Strong agent-oriented training: The models are fine-tuned for multi-turn dialogue, tool calling, and long-context reasoning — exactly what OpenClaw agents need.
2. Flexible sizing: Choose 27B for lightweight tasks, 35B-A3B for balanced workloads, 122B-A10B for advanced reasoning, or 397B-A17B when you need maximum intelligence.
3. Cost efficiency: Novita AI’s pricing for Qwen 3.5 is significantly lower than comparable models, making it practical to run 24/7 agents at scale.

Qwen 3.5 vs. Other Agent Models

Model	Best For	Typical Use in OpenClaw
Qwen 3.5 27B	General reasoning, task automation	Heartbeat checks, file management, message routing
Qwen 3.5 35B-A3B	Complex workflows, multi-step planning	Code review, document generation, research tasks
Qwen 3.5 122B-A10B	Advanced reasoning, long-context analysis	Deep research, complex debugging, strategic planning
Qwen 3.5 397B-A17B	Maximum intelligence, comprehensive analysis	Multi-domain research, enterprise-grade decision support
MiniMax M2.5	Balanced cost/performance	Good default choice
Kimi K2.5	Multimodal tasks	Screenshot analysis, UI automation
GLM-5	Long-context reasoning	Log analysis, large document processing

Pricing on Novita AI (Per 1M Tokens)

Model	Input Cost	Output Cost
Qwen 3.5 27B	$0.30	$2.40
Qwen 3.5 35B-A3B	$0.25	$2.00
Qwen 3.5 122B-A10B	$0.40	$3.20
Qwen 3.5 397B-A17B	$0.60	$3.60

How to Use Qwen 3.5 Series in OpenClaw (6 Steps)

Preparation

Before starting, ensure you have:

• A Linux, macOS, or Windows (WSL2/PowerShell) environment with internet access
• Node.js 22+ installed
• An active Novita AI account with API key

Step 1: Get Your Novita AI API Key

1. Visit Novita AI and create an account (or log in)
2. Navigate to API Keys in your dashboard
3. Click Add new key and copy it immediately (it’s only shown once)
4. Store the key securely — you’ll need it in Step 4

Step 2: Install OpenClaw

Run the one-line installer for your platform:

macOS / Linux / WSL2:

curl -fsSL https://openclaw.ai/install.sh | bash

Windows (PowerShell):

irm https://openclaw.ai/install.ps1 | iex

The installer will:

• Download and configure the OpenClaw CLI
• Install required dependencies
• Set up the Gateway service

Wait for the installation to complete, then verify:

openclaw --version

Step 3: Run the OpenClaw Configuration Wizard

Start the interactive setup:

openclaw config

You’ll configure four main sections: Workspace, Model, Gateway, and Channels. We’ll walk through each below.

3.1 Workspace Configuration

• Where will the Gateway run? → Select Local (for most users)
• Select sections to configure → Choose Workspace

This sets up your local environment for agent execution.

3.2 Model Configuration

• Select sections to configure → Choose Model
• Model/auth provider → Select Skip for now
• Filter models by provider → Choose All providers
• Default model → Select Enter model manually
• Default model → Enter: novita/qwen/qwen3.5-122b-a10b

Understanding model identifiers:

• Novita Model ID (on Novita platform): qwen/qwen3.5-122b-a10b
• OpenClaw format (requires provider prefix): novita/qwen/qwen3.5-122b-a10b

When configuring OpenClaw, always use the format novita/<model-id>. For Qwen 3.5 Series:

• novita/qwen/qwen3.5-27b
• novita/qwen/qwen3.5-35b-a3b
• novita/qwen/qwen3.5-122b-a10b
• novita/qwen/qwen3.5-397b-a17b

3.3 Gateway Configuration

• Gateway port → Default: 18789 (or customize if needed)
• Gateway bind mode → Select Loopback (Local only) for security
• Gateway auth → Choose Token authentication
• Tailscale exposure → Select Off (unless you need remote access)
• Gateway token source → Choose Generate/store plaintext token
• Gateway token → Leave blank to auto-generate (or set a custom token)

3.4 Channels (Optional: Telegram Example)

If you want to deploy your agent to a messaging platform:

• Select sections to configure → Choose Channels
• Configure/Link → Select Telegram (Bot API)
• Telegram Bot Token → Enter your bot token from @BotFather
• Choose DM access policy Select Pairing (users must pair before chatting)

Step 4: Configure Novita AI API Integration

The wizard doesn’t auto-configure Novita API credentials, so we’ll add them manually.

Open your OpenClaw config file:

open ~/.openclaw/openclaw.json

Add this block under "models" . Only update one field: <Your API Key> → your Novita API key

 "models": {
    "providers": {
      "novita": {
        "baseUrl": "https://api.novita.ai/openai/v1",
        "apiKey": "<Your API Key>",
        "api": "openai-completions",
        "models": [
          {
            "id": "qwen/qwen3.5-122b-a10b",
            "name": "Qwen3.5-122B (Novita)",
            "reasoning": false,
            "input": [
              "text"
            ],
            "cost": {
              "input": 0,
              "output": 0,
              "cacheRead": 0,
              "cacheWrite": 0
            },
            "contextWindow": 262144,
            "maxTokens": 65536
          }
        ]
      }
    }
  },

Step 5: Start the OpenClaw Gateway

Install and launch the Gateway service:

openclaw gateway install
openclaw gateway restart

Verify the Gateway is running:

openclaw gateway status

Step 6: Pair Your Messaging Platform (Telegram Example)

If you configured Telegram in Step 3.4:

1. Open your Telegram bot and send /start
2. The bot will reply with a pairing code (e.g., ABC123)
3. Approve the pairing in your terminal:

openclaw pairing approve --channel telegram ABC123

Your Telegram account is now paired! Send a message to test:

Hello! What's the weather today?

The agent will respond using Qwen 3.5 122B-A10B (your default model).

Optimizing Token Usage: Choosing the Right Model for Each Task

One of Qwen 3.5 Series’ biggest advantages in OpenClaw is flexible sizing. Instead of using the largest model for everything, route tasks intelligently:

Recommended Model Selection

Task Type	Recommended Model	Why
Heartbeat checks, status updates	Qwen 3.5 27B	Fast, cheap, good enough for simple queries
File summaries, web searches	Qwen 3.5 27B or 35B-A3B	Balanced reasoning without overkill
Code generation, multi-step workflows	Qwen 3.5 35B-A3B or 122B-A10B	Strong instruction following + tool use
Complex debugging, research synthesis	Qwen 3.5 122B-A10B	High-level reasoning depth + long-context support
Enterprise decision support, multi-domain research	Qwen 3.5 397B-A17B	Maximum intelligence for mission-critical tasks

Conclusion

OpenClaw’s explosive growth in 2026 has proven the demand for flexible, powerful AI agent frameworks — but unchecked token consumption can make deployment unsustainable. By pairing OpenClaw with Qwen 3.5 Series on Novita AI, developers gain:

• Cost efficiency: Up to 80% savings vs. premium models, with flexible sizing (27B → 397B)
• Production-ready performance: Strong reasoning, tool use, and long-context support
• Scalability: Run 24/7 agents without burning budgets

Whether you’re building customer support bots, internal automation workflows, or research assistants, this combination delivers enterprise-grade capabilities at startup-friendly prices.

Ready to deploy? Grab your Novita AI API key and start building smarter, cheaper agents today.

Novita AI is a leading AI cloud platform that provides developers with easy-to-use APIs and affordable, reliable GPU infrastructure for building and scaling AI application.

Frequently Asked Questions