TL;DR
  • Tested 4 AI models for WhatsApp group @mention capability
  • Tested: Claude Sonnet 4.6 is the only model that executes @mention 100% perfectly
  • Free models (Qwen, MiMo) need server-side patches and have unstable accuracy
  • Key finding: AI's instruction-following ability directly determines whether mention works

Background

When an OpenClaw AI assistant @mentions other members in a WhatsApp group, the mentionedJid field in the message must correctly contain the LID (Linked ID) for WhatsApp to display it as a blue clickable mention and trigger a notification.

We developed a custom patch in OpenClaw that automatically converts @Name to WhatsApp LID mention before sending the message. However, whether the patch works depends on whether the AI model outputs the correct format.

Test environment: OpenClaw v2026.3.28 + WhatsApp Web

Test Models

Model Provider Price Type
Claude Sonnet 4.6 Anthropic Paid Closed-source
Qwen 3.6 Plus Preview OpenRouter Free Open-source
Xiaomi MiMo v2 Pro OpenRouter Free Open-source
StepFun Step 3.5 Flash OpenRouter Free Open-source

Test Results

Claude Sonnet 4.6 (Anthropic)

PERFECT
Behavior Precisely follows SOUL.md instructions, directly outputs LID number format (e.g. @275806039314634)
WhatsApp Blue clickable mention -- no patch needed
Instruction Following Excellent -- can read group LID mapping table and use correctly
Notes The only model that perfectly executes @mention

Qwen 3.6 Plus Preview (OpenRouter, Free)

Needs Patch
Behavior Outputs @Justin (name format), relies on patch to convert to LID
WhatsApp Can display as real mention after patch conversion
Instruction Following Medium -- understands @ concept but doesn't follow LID format instructions
Notes Requires server-side patch support to work

Xiaomi MiMo v2 Pro (OpenRouter, Free)

Wrong Target
Behavior Can trigger blue mention correctly, but sometimes @s wrong person
WhatsApp Blue clickable mention -- but may target wrong person
Instruction Following Below average -- understands @ but not precise
Notes Mention mechanism works, but accuracy insufficient

StepFun Step 3.5 Flash (OpenRouter, Free)

Failed
Behavior Doesn't output @ format at all, only replies in plain text
WhatsApp No mention
Instruction Following Poor -- doesn't understand group @mention concept
Notes Not suitable for scenarios requiring mention functionality

Overview Comparison

Dimension Claude Qwen MiMo StepFun
@Mention Perfect Needs Patch Wrong Target Failed
Instruction Following 5/5 3/5 2/5 1/5
Value Paid Free Free Free
Recommended Use Commercial / Critical Daily Use Experimental Not Recommended

Technical Architecture

The following is the complete data flow of WhatsApp @mention from user input to final display:

// WhatsApp @Mention Data Flow

User @Bot
    |
    v
WhatsApp -------> OpenClaw Gateway -------> AI Model
                                                  |
                                                  v
                                        Reply text with @Name
                                                  |
                                                  v
                                   deliver-reply patch
                                   @Justin -> @275806039314634
                                                  |
                                                  v
                                        reply() function
                                   Detect numeric LID
                                   Add to mentions[]
                                                  |
                                                  v
                                   WhatsApp renders blue @mention

Key patch files:

File Function
deliver-reply-CCZOVb0X.js Converts @Name to @LID number before sending
login-B5O9Mtcp.js reply() Detects @numeric LID and adds to WhatsApp mentions array
login-B5O9Mtcp.js sendTrackedMessage() Mention injection before final send
send-DtEayToE.js Mention injection for message tool path

LID name mappings are stored in /home/openclaw/.openclaw/workspace/LID_CACHE.json, automatically cached from group metadata.

Key Findings

The AI model's "instruction-following ability" directly determines whether WhatsApp @mention functionality works.

Even with a well-implemented underlying patch, weak models still cannot correctly generate mention text. When choosing an AI model, instruction-following ability should be a key evaluation criterion.

The test results of the four models show a clear gradient: from Claude's perfect execution, to Qwen's "understands but not precise," to MiMo's "triggers but inaccurate," and finally StepFun's complete lack of understanding. This gradient essentially reflects the differences in each model's ability to follow complex, structured instructions.

For commercial deployment or mission-critical scenarios, Claude remains the only reliable choice. Free models can achieve basic functionality with patch assistance, but stability and accuracy remain bottlenecks.

About InstallMyClaw

InstallMyClaw is Malaysia and Singapore's #1 OpenClaw installation service. We help you install your AI assistant on Mac, Windows, or Linux and connect it to WhatsApp, Telegram, and Discord. Whether you're setting up an AI WhatsApp bot or deploying OpenClaw on a VPS — we handle everything remotely in under 4 hours. DIY is free, or choose managed subscriptions from RM99/mo (MY) / SGD35/mo (SG).

WhatsApp Us