What is WeLM
WeLM (WeChat Language Model) is a large language model independently developed by WeChat and has been iterated to V4. The model is purpose-built for the WeChat ecosystem, with approximately 80B-130B parameters, adopts a sparse MoE architecture, and emphasizes low cost and high efficiency. WeLM supports a 128K long context and can naturally access ecosystem data such as group chats and Moments. Through Hidden Decoding technology, it hides the reasoning process to achieve fast responses. WeLM is not open to the public and is the intelligent foundation of WeChat AI Agent Xiaowei.
WeLM's main features
- Exclusive intelligent services for the WeChat ecosystem: As the underlying model of the WeChat AI Agent "Xiaowei", WeLM is deeply embedded in the WeChat ecosystem and supports scenarios such as group chat message summarization, Moments feed viewing, friend reply phrasing, and understanding of Official Account/Channels content.
- Low-cost and efficient inference: It adopts a highly sparse MoE (Mixture of Experts) architecture, combined with technologies such as GQA, KV-Mirror, and Multi-Token Prediction, greatly reducing inference costs and supporting the high-frequency calls of WeChat's 1 billion+ daily active users.
- Fast response (Hidden Decoding): Through Hidden Decoding technology, the reasoning process is hidden, achieving low-latency responses while ensuring output quality and avoiding long waits for users.
- Long-context understanding: It supports an ultra-long 128K context and naturally accesses ecosystem data such as group chat records, Moments, favorites, and likes, enabling precise intent recognition and personalized replies.
- Multi-task processing: It covers multiple tasks such as content summarization, dialogue generation, information retrieval, and personalized recommendations, becoming the core engine of AI capabilities within WeChat.
How to use WeLM
- Visit the official entry point: Search for "Xiaowei" in WeChat to find the entry point.
-
Update WeChat: Upgrade WeChat to the latest version to ensure you can receive Xiaowei beta pushes or full rollout access.
-
Find the entry point: Find the "Xiaowei" AI Agent entry in the WeChat chat list, top search bar, or Discover page.
-
Start a conversation: Tap "Xiaowei" and directly enter text or voice to submit your questions and task requirements.
-
Authorize data: Follow the prompts to authorize access to ecosystem data such as group chat records, Moments, and favorites to get personalized, context-aware replies.
-
Execute tasks: Use "Xiaowei" for operations such as group chat summarization, content recommendations, phrasing generation, and information retrieval.
-
Optimize interaction: Based on Xiaowei's feedback, continuously refine your instructions through multi-turn conversations to improve result quality.
WeLM's core advantages
- Native integration with the WeChat ecosystem: WeLM is a "dedicated small model" purpose-built for the WeChat ecosystem. It naturally accesses contextual data such as group chats, Moments, Official Accounts, Channels, and favorites, without requiring users to manually upload or authorize external tools, enabling it to understand user intent more precisely.
- Ultimate cost control: It adopts a highly sparse MoE architecture (not an ultra-large dense model), combined with technologies such as GQA, KV-Mirror, and Multi-Token Prediction, greatly reducing inference costs and supporting the high-frequency calls of WeChat's 1 billion+ daily active users.
- Fast response experience: Using Hidden Decoding technology to hide the reasoning process, achieving low latency while ensuring output quality, avoiding user waiting, and adapting to WeChat instant messaging scenarios.
- Ultra-long context memory: Supports 128K long context, performing excellently in long-text tasks such as group chat summarization, Moments review, and personalized recommendations.
- Stable and controllable closed architecture: Not open to the public, focused on serving internal WeChat scenarios, with inherent advantages in stability, security, and ecosystem consistency.
Comparison of WeLM's similar competitors
| Dimension | WeLM | ByteDanceDoubao |
|---|---|---|
| Positioning | Closed model dedicated to the WeChat ecosystem | General-purpose AI assistant for the ByteDance ecosystem (open across all platforms) |
| Parameter scale | 80B-130B ("small model") | Hundreds-of-billions-scale general large model |
| Architecture | Highly sparse MoE, extreme cost compression | Dense + MoE hybrid, pursuing general performance |
| Data integration | Native access to WeChat ecosystem data (group chats, Moments, etc.) | Requires users to manually upload or authorize external data |
| Response strategy | Hidden Decoding hides reasoning, pursuing extreme speed | Supports displaying the reasoning process, enabling deep thinking |
| Degree of openness | Not open to the public, serving only internal WeChat | Fully open via standalone app, web, and API |
| Core goals | Low cost, low latency, high stability, deep ecosystem | General capabilities, multimodality, large-scale user coverage |
WeLM's application scenarios
- Intelligent group chat summarization: WeLM can automatically sort through massive historical messages in WeChat group chats, accurately extracting key information, core decisions, and to-do items, helping users quickly grasp group chat dynamics.
- Moments interaction assistance: Based on the content of friends' Moments posts and combined with the relationship between the user and the friend, WeLM can intelligently generate appropriate, personalized comment or reply suggestions.
- Ecosystem content Q&A: WeLM can provide in-depth interpretation of ecosystem content such as WeChat Official Account articles and Video Accounts, instantly answering users' questions about article details, video topics, and more.
- Personal memory retrieval: WeLM supports personal data such as cross-chat records, favorites, and like history, helping users quickly and accurately locate and recall important past information.
- Intelligent phrasing generation: WeLM can combine the current chat context, friend relationships, and personal style to recommend appropriate, natural reply wording and communication phrasing for users.