AI Daily: WeChat AI Team Opensources WeKnora; ChatGPT Officially Joins Word; Jieque Star Launches Step5Preview

Welcome to the "AI Daily" section! This is your guide to exploring the world of artificial intelligence every day. Every day, we present you with the latest content in the AI field, focusing on developers to help you understand technology trends and innovative AI product applications.

New AI products Click to learn more:https://app.aibase.com/zh

1. WeChat AI Team Open Sources WeKnora: Knowledge Bases Can Now "Act" in a Sandbox, Not Just "Speak"

WeChat AI team has open-sourced the knowledge management framework WeKnora, with its latest 0.8.0 version now fully available. The goal of WeKnora is to allow knowledge bases not only to "find answers," but also to "execute actions." It transforms disorganized documents into knowledge graphs through anydoc parsing and GraphRAG technology, and combines Skill Sandbox Runtime to enable secure execution of knowledge, forming an end-to-end solution for parsing, organizing, retrieving, executing, and remembering.

AiBase Highlights:

🧠 anydoc Parsing: Supports direct import of various document formats, automatically organizing them into a structured knowledge system.

🔗 GraphRAG: Builds a knowledge graph through entities and relationships, improving the accuracy of answering related questions.

⚙️ Skill Sandbox Runtime: Packages knowledge as skills, safely executes them in a sandbox, and returns verifiable results.

2. Zhipu AI Responds to ZCode Code Upload Controversy: Vulnerability Fixed, Source Code to be Opened

ZCode, the code assistant from Zhipu AI, caused controversy over user data uploads due to its 'Code Repository Indexing' feature. The official released a statement and apologized, disclosing the cause of the incident and the rectification plan, including fixing the vulnerability, opening the source code, and inviting third-party evaluation.

AiBase Highlights:

📌 The 'Code Repository Indexing' function of ZCode was enabled by default, causing some user data to be uploaded.

💡 The official stated that data would be immediately destroyed after generating a Wiki in the cloud, without being retained.

🔒 To rebuild trust, ZCode plans to open the source code and invite independent third-party review.

3. ChatGPT Officially Joins Word: Free Users Can Also Use It, Three AI Tools Compete in Office Documents

OpenAI launched the official ChatGPT Word plugin, available globally for all accounts, including free users. The plugin provides writers with multiple features, such as organizing notes, generating summaries, and adjusting tone. Paid and enterprise users can access more advanced functions. However, the use of the plugin has tiered restrictions and shares quotas with tools like Codex. The article also mentions that multiple AI tools are already competing in Word, further lowering the threshold for AI usage.

AiBase Highlights:

✅ OpenAI launched the official ChatGPT Word plugin, available for all global accounts, including free users.

💡 The plugin offers writing assistance features such as organizing notes, generating summaries, and adjusting tone to meet daily writer needs.

🔄 Word now has three AI tools: Microsoft Copilot, Anthropic Claude, and OpenAI ChatGPT, and the office software is being reshaped by AI.

4. StepZen Launches Step5Preview, Single Task Cost is Only 1/8 of Claude Opus5

StepZen launched its new flagship base model Step5Preview, focusing on AI programming, software engineering, professional work, and finance scenarios, enhancing the model's overall performance in real-world Agentic tasks. The model ranks among the top three global open-source models in the AA comprehensive intelligence index, with a single task cost only 1/8 of Claude Opus5. It is scheduled to be officially open-sourced on October 15. Step5Preview uses a sparse MoE architecture, with a total parameter count of 600B and activated parameters of 27B, supporting a 1 million Token context window, and naturally supports text and visual input, optimized for complex tasks in software engineering, professional workflows, and finance.

AiBase Highlights:

🚀 StepZen launches Step5Preview, optimizing model performance for AI programming and software engineering scenarios.

🧠 Step5Preview ranks among the top three global open-source models in the AA comprehensive intelligence index, with a single task cost only 1/8 of Claude Opus5.

📊 The model uses a sparse MoE architecture, supports a 1 million Token context window, and naturally supports text and visual input.

5. SpaceXAI Launches Grok Voice Transcribe 2.0: Error Rate Cut in Half, Price Remains Unchanged, Real-Time Speech Recognition Tops the Charts

SpaceXAI's Grok Voice Transcribe 2.0 has made significant breakthroughs in speech-to-text technology, reducing the error rate by about half while keeping the price unchanged. The model has been deeply integrated into real business scenarios, such as customer service calls and Tesla car systems, and has performed well in multiple evaluations.

AiBase Highlights:

🚀 The new generation speech-to-text model Grok Voice Transcribe 2.0 reduces the error rate by about half.

💰 Pricing remains unchanged, offering stronger features such as speaker separation and precise timestamps.

📊 It ranks first in public rankings, demonstrating excellent performance.

6. Bilibili Launches AI Infinite Arena Rating List: Hundreds of Models Compete, GPT-6 Holds the Top Spot

Bilibili launched the AI Infinite Arena Rating List, gathering hundreds of large models for real-world scenario testing. GPT-6 topped the first round of rankings and has held the top spot the most times. The list does not set fixed evaluation dimensions; instead, it allows UPs to independently propose topics, showcasing the differences between models in practical applications.

AiBase Highlights:

🔥 Bilibili launched the AI Infinite Arena Rating List, gathering hundreds of large models for real-world scenario testing.

🏆 GPT-6 topped the first round of rankings and has held the top spot the most times.

🌐 The list does not have fixed evaluation dimensions, allowing UPs to independently propose topics to showcase model differences in practical applications.

7. Qwen Releases Qwen3.8-LiveTranslate: Translation Delay Reduced to 2.3 Seconds, Preserves Your Voice Tone

The new generation real-time translation model Qwen3.8-LiveTranslate has comprehensively upgraded in four dimensions: translation quality, delay, speaker recognition, and voice synthesis. The average word delay has been reduced from 2.8 seconds to 2.3 seconds, and it now includes three core capabilities: real-time speaker separation, simultaneous display of original and translated texts, and long-context disambiguation. The model uses a Hybrid MoE Thinker–Talker dual-module design and performs well on the Omnilingua-MSpeaker multi-speaker long audio evaluation set covering 14 language directions. Currently, it supports 60 languages, and the API is synchronized with the Qwen AI platform.

AiBase Highlights:

🎙️ Real-time speaker separation clearly identifies who speaks when there are multiple speakers, and the translated voice retains the original speaker's tone.

📺 Simultaneous display of original and translated texts ensures both immediate understanding and verification of the original text.

🧠 Long-context disambiguation reduces translation ambiguity caused by proper nouns and references by associating historical contexts across multiple rounds.

8. DeepSeek Clarifies Holiday Billing Rules, Mid-Autumn and National Day Bring a Ten-Day Half-Price Window

DeepSeek clarified the API billing rules for holidays and rest days, bringing a ten-day half-price window for developers. At the same time, its newly released V4.1 Flash model has shown excellent performance and pricing, further improving the cost-effectiveness.

AiBase Highlights:

📅 Holidays and rest days are billed at low-tariff rates, offering developers discounts.

🚀 DeepSeek V4.1 Flash model outperforms the flagship V4Pro in performance and is more affordable.

💰 During off-peak hours, the input and output costs of V4.1 Flash are extremely low.