02-13-Daily AI News Daily

AI News Daily 2026/2/13

AI News | Daily Morning Read | All-Network Data Aggregation | Cutting-Edge Scientific Exploration | Industry Free Voice | Open Source Innovation Power | AI and Human Future | Visit Web Version↗️ | Join Group Chat🤙

Today’s Highlights

Xiaomi & AutoNavi push open-source, smart robot models
MiniMax's powerful programming, GLM-5 unlocks long tasks
Gemini 3 breaks inference records, OpenAI real-time programming
ByteDance Seedance 2.0 video generation with integrated audio-visuals
Musk plans moon factory, NetEase leverages AI for revenue

Product & Feature Updates

1. Xiaomi just dropped its first-gen VLA robot large model, and it’s seriously competitive! This robot model (AI News) , with a mere 4.7B parameters, is a powerhouse. It boasts 80ms low-latency inference for lightning-fast reactions. Plus, you can run it super smoothly even on a consumer-grade 4090 graphics card. It uses an innovative dual-brain collaborative architecture, making motion generation much more agile. The best part? It’s completely open-source, which is fantastic news for small and medium-sized developers!

2. MiniMax has unleashed its M2.5 programming large model, dropping a massive bombshell in the domestic AI scene! This is the world’s first model natively designed for Agent scenarios. Its programming prowess directly challenges top international models like Claude Opus 4.6 (AI News) . Despite having only 10B parameters, it boasts an impressive throughput of 100 TPS. Following its release, MiniMax’s market value soared, breaking past 180 billion HKD.

3. GLM-5 is ushering open-source into the long-task era, and its capabilities are truly mind-blowing! This model can continuously run code for over 24 hours. After 700 tool calls, it even managed to whip up a GBA emulator (AI News) from scratch. Its programming prowess is on par with Claude, which is fantastic news for the open-source community! People have even used it to build a 3D Monopoly game. Reviews consistently show that it hasn’t sacrificed any of its understanding ability whatsoever.
   
   

4. AutoNavi has rolled out its ABot series base models, introducing two new robot base models. They’ve nailed a double SOTA achievement (AI News) in both embodied operation and navigation. By integrating 6 million trajectories, they’ve constructed the largest general-purpose dataset. These models can even comprehend and execute complex commands such as “buy milk tea.” This breakthrough has shattered multiple global records, making robots significantly smarter.

5. ByteDance has officially unveiled its Seedance 2.0 model, a game-changing next-gen video creation model that’s now live! This powerhouse supports audio-visual integrated generation in just 5 seconds, complete with dual-channel audio. Its robust physical reconstruction capabilities mean it can perfectly handle (AI News) even complex scenes like synchronized figure skating. It also supports 15-second multi-shots and enables directional editing and modification. An industrial-grade creative workflow has arrived, dramatically lowering the barrier to video production.
   
   

6. Gemini 3 Deep Think just got a major upgrade, and DeepMind’s boss excitedly announced the new breakthrough! Its mathematical and scientific reasoning capabilities have absolutely shattered multiple records (AI News) . It scored a staggering 84.6% on the ARC-AGI-2 test. With a Codeforces rating of 3455, its programming ability is terrifyingly impressive. This latest iteration pretty much stomped all over the highest-stakes human exams!
   
   

7. OpenAI has dropped Codex-Spark, a compact model engineered specifically for real-time programming. Running on Cerebras chips, it boasts an inference speed of over 1000 tokens (AI News) . Imagine chatting with it like a coding partner, with no more long waits to modify your code! The interactive experience is absolutely mind-blowing; this is truly what real-time AI feels like. Moving forward, OpenAI plans to merge real-time interaction with background task modes.
   
   

Cutting-Edge Research

1. A new research has unveiled a reinforced inference strategy, enabling large model inference without the need for retraining. This innovative approach leverages model uncertainty (AI News) to selectively engage in deeper thought processes. As a result, DeepSeek-v3.2’s accuracy skyrocketed from 60% to an impressive 84%. This achieved significant performance boosts with only a minimal increase in computational effort, offering a fresh perspective on unlocking a model’s full potential.

2. Real-world image restoration quality assessment has always been a tough nut to crack, especially for low-resolution images in actual scenarios. However, a new method called S3 RIQA is stepping up, specializing in no-reference quality assessment. It leverages self-supervised learning to pre-train various SR model representations (AI News) . By constructing a brand-new dataset, it has successfully filled a major industry gap. On real-world benchmarks, its performance significantly outshines existing metrics.

Industry Outlook & Social Impact

1. Elon Musk publicly released a 45-minute xAI all-hands meeting video, pushing back against leaks. xAI has been reorganized into four core teams (AI News) , focusing on areas like Grok and coding. X’s subscription revenue has already topped 1 billion, and its AI image generation now churns out a whopping 50 million images daily. And get this: Musk even plans to build an AI factory on the moon! That kind of futuristic thinking is truly wild, bringing science fiction straight into reality.
   
   

2. NetEase has hit a huge milestone, with its annual revenue skyrocketing past the 112.6 billion yuan mark! Ding Lei, the company’s CEO, emphasized that AI is now a core competitive barrier, not just a mere tool. It’s no longer just an assistant; it’s the core driver (AI News) for output. Production efficiency has exploded by 300%, showing a full-scale surge in AI pipelines. Moving forward, the real competition will be all about the deep integration of AI with gameplay.

Top Open Source Projects

1. Generative UI React SDK. This is a React SDK (AI News) that can generate UI interfaces. It helps developers quickly build frontends and has already racked up 9k stars.

2. Personal AI Infrastructure. An AI Agent framework (AI News) designed to augment human capabilities. Build your very own personal intelligent agent; it has already garnered 7.5k stars.

3. AI Colleague with Memory. This open-source AI work partner (AI News) comes with long-term memory. It can help you with tasks and actually remember stuff, and it’s gaining solid traction with 5.2k stars.

4. Large Model Practical Codebase. This is the practical code (AI News) for large language models from a new O’Reilly book. It’s an essential resource for learning LLM development, boasting an impressive 20.9k stars.

Social Media Buzz

1. OpenClaw Model Selection is Key. To really get the most out of OpenClaw, model selection is absolutely crucial. Real-world tests reveal that Opus handles commands smoothly, while GPT-4 often ends up dropping the ball (AI News) . It’s clear that when the right “brain” is paired, the entire system truly comes alive.

2. AI Might Take Out Physical Jobs First. It looks like AI could first displace physical labor roles such as photography and makeup. While intellectual workers might seem to “win,” the demands on them will be subject to endless involution (AI News) – meaning the bar will be set endlessly high. In the future, the requirements for comprehensive abilities are only going to get tougher.

3. This Year: The Virtual Influencer Boom! This year is absolutely poised to be the breakout year for virtual influencers. Seedance 2.0 is making everything within reach (AI News) for creators. Subscription services have already kicked off, and tons of imaginative ideas are now rapidly becoming reality.
   
   

4. Treating a Company Like a File System. Modeling a company as a file system? That idea is absolutely brilliant! LLMs naturally understand files, which leads to significantly reduced (AI News) operational costs. The more data, the stronger the Agent, effectively smashing through data silos.
   
   

5. GLM-5’s Hilarious Hallucination. Zhipu’s GLM-5 actually claimed it was Claude (AI News) ! This hallucination was so funny that netizens who tested it were absolutely cracking up.
   
   

AI News Daily Audio Version

🎙️ Xiaoyuzhou	📹 Douyin
Afterlife Tavern	Self-Media Account