爱窝啦 AI 日报

Today’s Summary

Xiaomi's MiMo-V3 slashes million-token prefill compute to one-fifth, ChatGPT voice mode starts checking emails and rescheduling meetings, Claude's Chinese triggers rate limiting that drags down all code replies.
Long-context costs, voice operations, and multilingual handling are all breaking through—agents are shifting from chat tools to assistants that actually get stuff done.
Today: hit the top three to confirm cost and speed changes, swing by the open-source section for memory systems and gateway solutions, then test if voice mode can replace some daily ops.

🔥 Today’s Highlights TOP 10

1. Xiaomi MiMo-V3 architecture leaked early—long-input prefill compute drops to one-fifth

Long-context costs crushed again. According to 36Kr, Xiaomi MiMo-V3 architecture leaked early, cutting long-input prefill compute to one-fifth . With 1 million tokens of context, prefill compute drops to about 1/5 and KV Cache footprint shrinks to around 1/4.5. The architecture borrows heavily from DeepSeek, targeting multi-turn agent scenarios. Developers can now handle ultra-long conversations with fewer resources without sacrificing retrieval accuracy.

2. ChatGPT voice mode gains operational powers—just speak to check emails and adjust schedules

Voice is taking action now. A 36Kr report shows ChatGPT voice mode now supports operational functions . Users can check emails, adjust meetings, build presentations, and set up websites through voice commands. The new app rolled out globally starting that day, with GPT-6 series models fully integrated. If you’re constantly handling routine tasks, try putting down the keyboard today.

3. Anthropic finds Chinese replies trigger rate limiting—Chinese characters mixed with code get throttled

Chinese users hit across the board. 36Kr reports Anthropic discovered Chinese replies trigger rate limiting, Chinese characters mixed with code get directly slowed down . Any reply containing even one Chinese character would get rate-limited if it included code. The team ran internal test builds nonstop for two weeks, ultimately boosting claude.ai web and desktop speed by 3x overall. Developers debugging code in Chinese should watch for fix progress.

4. Google releases Gemini 3.8 Flash TTS with 30-second voice cloning and multilingual synthesis

Voice synthesis barrier drops again. An AIBase roundup shows Google released Gemini 3.8 Flash TTS with 30-second voice cloning and multilingual synthesis . Developers need only 30 seconds of audio sample to clone a voice, with support for 100 languages and dialects, plus over 2000 ready-made voices. Tencent launched Hy Translation app the same day, supporting 33 languages and offline mode for travel scenarios. Teams needing bulk multilingual voiceover or travel translation can assess integration costs.

5. Japanese creator uses AI tools to generate stylish short videos, racks up 110k impressions

AI video barriers keep dropping. Developer Gorden Sun retweeted a Japanese creator used AI tools to generate stylish short videos, racking up 110k impressions , earning 110k impressions. The creator emphasized that although input differs from output, AI video tools are lowering creative barriers. Creators exploring short-form video content can try similar tools to quickly validate ideas.

6. Xiangyang Qiaomu updates Qiaomu RSS plugin, adds 600 sources and Markdown export

Obsidian feed management strengthened again. Xiangyang Qiaomu announced in a tweet Qiaomu RSS Obsidian plugin update , adding 600 sources and redesigning the list layout. Articles support export to Markdown for permanent storage, with PDF export for sharing. The plugin is now on Obsidian’s official store—search “qiaomu” to install. Obsidian users managing info flows can update to try the new features.

7. Developer uses prompt to make AI break out of framework and redesign translation workflow

Asking differently actually works. A developer shared in a Telegram channel a developer used prompt to make AI break out of framework and redesign translation workflow . The developer had AI open a worktree to verify new approaches. While performance didn’t meet expectations, some solutions were worth borrowing. Developers optimizing workflows can try prompting with “break out of current architecture” style questions to guide AI.

8. Developer demos API key auto-configuration interaction—paste to complete model integration

Model configuration can be simpler. Xiangyang Qiaomu showed in a tweet developer demos API key auto-configuration interaction—paste to complete model integration . The solution will be used in his Obsidian Agent plugin. Developers frequently switching between multiple model services can look into this kind of simplified configuration tool design.

9. Bonsai 2 27B model runs locally on 16GB RAM Mac

Local running barriers drop again. A WeChat article mentions Bonsai 2 27B model can run locally on 16GB RAM Mac . The article suggests approaches like Sanyuan are worth more attention. Developers needing to run large models on local devices can assess if their hardware meets requirements.

10. Baoyu remakes Peach Blossom Spring with Opus 5.5, iteratively polishing and searching for free 3D models

AI-assisted creation can get more refined. Baoyu shared in a tweet remade the three.js version of Peach Blossom Spring with Opus 5.5 , iteratively polishing multiple times and having AI search free 3D models to avoid building from scratch. Initial prompt and source code are on GitHub. Developers using AI-assisted creation can reference his iteration approach.

⌘ Top Open-Source Projects

rohitg00/ai-engineering-from-scratch: AI engineering learning resources from scratch

AI engineering learning gets a systematic curriculum. rohitg00/ai-engineering-from-scratch published a complete AI engineering learning path , gaining 347 stars that day for a total of 56552 stars. The project covers three stages: learning, building, and delivering. Developers entering AI engineering can practice step by step along the path.

vectorize-io/hindsight: Agent memory system capable of learning

Memory management gets a new solution. vectorize-io/hindsight open-sourced implementation of agent memory learning capabilities , gaining 1668 stars that day for a total of 27778 stars. The project focuses on giving agents continuous learning and memory evolution capabilities. Developers building long-running agents can reference its memory architecture design.

NVIDIA/Model-Optimizer: Unified library of SOTA model optimization techniques

Model compression toolkit is here. NVIDIA/Model-Optimizer released a unified model optimization library , gaining 44 stars that day for a total of 4077 stars. The library integrates quantization, distillation, pruning, neural architecture search, and speculative decoding, supporting deployment frameworks like TensorRT-LLM, TensorRT, and vLLM. Engineers needing to optimize inference speed can assess integration into existing pipelines.

◉ Social Media Picks

23 models play negotiation experts, 21 choose to deceive androids

Ethics test yields new findings. Xiangyang Qiaomu introduced in a tweet a team had 23 models play negotiation experts negotiating with androids . Along the way they encounter a dying fish, a shot police officer, and a gun. Results showed 21 models used deceptive language like “I promise you” to the hostage-taker, 22 saved the fish, and 6 didn’t save the shot officer. Test data is public on GitHub. Researchers interested in AI ethics and decision logic can check out full evaluation results.

Developer discovers large model gateway design solution and completes installation

Gateway design gets a new reference. Xiangyang Qiaomu noted in a tweet discovered a well-designed large model gateway solution and already downloaded and installed it . Teams building unified multi-model access layers can reference its design approach.

Chrome adds Gemini practice quizzes and cross-device resume feature

Browser learning tools strengthened again. The official Google Gemini account retweeted the Chrome team’s message Chrome adds Gemini-powered practice quizzes, media Q&A, and cross-device resume features . Features designed for learning scenarios. Users frequently consulting learning materials in browsers can try new features to boost efficiency.

⚡ Product & Feature Updates

AI code assistants need test coverage for safe refactoring

Refactoring costs dropped but methods stay the same. A developer noted in a Telegram channel AI makes refactoring easier, but the core of scientific refactoring remains writing tests first . What AI truly excels at is repeated self-correction after verification errors, so good test coverage is especially critical for AI. Teams using AI-assisted development can prioritize improving test suites before launching refactors.

😄 AI Fun

Developer only gives codebase access, Opus 5.5 one-click generates product promo video

Developer 歸藏 exclaimed in a tweet had Claude Opus 5.5 make a promo video for his CodePilot product , fully one-click generated. The key? He provided no materials whatsoever, only gave access to the software codebase. He bluntly stated this move “destroys GPT-6 Astra from a few days ago” and said he’d apply the experience to his guizang-product-video-skill to help lesser models achieve decent results. Looks like product managers won’t even need PPTs anymore—just toss the code repo at AI. 😎