爱窝啦 AI 日报

Today’s Summary

DeepSeek drops desktop source code, Ant-Tsinghua's 9B model runs dialogue and tasks in parallel, Qualcomm goes all-in on agent platforms
Opus 5.5 extends context, GPT-6 optimizes caching, training costs drop to $100-level—everyone's fighting for the same thing: getting models into real workflows
Check out DeepSeek's desktop client and Together's training tutorial today, then decide whether to migrate your debug environment or run a low-cost training experiment

🔥 Today’s Top 10 Highlights

1. DeepSeek Harness Releases Desktop Preview

Holiday update dropped. According to 36Kr, DeepSeek Harness releases desktop preview , with official code explicitly pointing to download.deepseek.com as the update source. The apps/desktop directory has had complete architecture since late August. Developers waiting for Harness local debugging features can try the desktop experience today.

2. Ant-Tsinghua Open-Sources 9B Real-Time Dialogue Model

Multi-turn conversations can now handle tasks asynchronously. According to QbitAI, Ant-Tsinghua open-sources 9B real-time dialogue model with 9B parameters. The model uses one Harness to connect frontend and backend, responding to follow-up questions while simultaneously fetching data or running tasks. Developers who need simultaneous dialogue and task execution can check out the repo’s implementation today.

3. Claude Opus 5.5 and GPT-6 Luna Bet on Different Tech Routes

Both hunting for new breakthroughs. According to 36Kr’s analysis, Claude Opus 5.5 and GPT-6 Luna bet on different technical approaches —the former borrows Hunyuan’s long-context design, the latter references DeepSeek’s caching optimization. Artificial Analysis data shows Opus 5.5 outputs an average of 119k tokens per query in max mode. Both new models attempt to reduce compute costs and boost intelligence through different technical paths.

4. Tencent AI Assistant QClaw Announces Shutdown

Another big-tech AI product exits. According to 36Kr, QClaw will cease service on December 24, 2026 at midnight , users can download backups until March 24, 2027, with official migration recommended to WorkBuddy. Shutdown reason cited as business restructuring and resource consolidation. Current QClaw users can start backing up data today and evaluate alternatives.

5. Chinese LLMs Enter Compute and Inference Resource Crunch

Old feathers shed, new ones not yet grown. A Jike user analyzes that Chinese LLMs are going through an awkward phase . Alibaba mentioned at Apsara Conference they might train models at 5T to 10T parameter scale, and Kimi, GLM are also scaling up. But training chip growth can’t keep pace with model scale, and inference resources also present real constraints. Teams planning model training need to reassess compute resources and inference costs today.

6. Meta Bets on Wearables as Next-Gen AI Entry Point

Zuckerberg’s real target isn’t phones. A Jike user analyzes that Meta bets on wearables as the next-gen AI entry point . Silicon Valley consensus sees phones as transitional, with the real next-gen platform being wearables like glasses, watches, earbuds, rings. But for hardware entry points to work, the ecosystem needs enough smart services. Meta missed operating systems in both PC and mobile internet eras—this time they want to position early.

7. Together AI Releases Low-Cost Model Training Tutorial

Train a Jev-like model for 100 RMB. XiangYangQiaoMu introduces in a tweet that Together AI releases low-cost model training tutorial , using 8 public datasets from Huggingface with 38k examples, taking 25 minutes at roughly $17 cost. Though it feels like platform promotion, this price point does let individual developers try training from scratch today.

8. Today AI Pushes Personalized News but Raises Cost Concerns

What’s the price of precision delivery? A Jike user shares that Today AI pushes personalized news but raises cost concerns . User thinks the design is good but worries backend costs are too high and public data quality is easily polluted. Pricing and information source credibility are issues current personalized AI news services need to solve.

9. Google Plans to Launch TPU-Equipped Satellite Next Week

AI chips going to space for inference. Baoyu introduces in a tweet that Google plans to launch TPU-equipped satellite next week , scheduled for October 1. The satellite carries 4 TPUs, with solar panels providing only about 1 kilowatt power. Chips run for 15 minutes then must cool down, with design lifespan about one year. Google wants to test whether AI data centers can move to space, leveraging near-Earth orbit solar power that can reach up to 8x ground levels.

10. Claude Subscription Now Recommended via US Apple Store

Nigeria region no longer cheap. Gorden Sun suggests in a tweet that Claude subscription now recommended via US Apple Store , as Nigeria region pricing advantage has faded. He shares subscription process screenshots and says accounts opened by him and friends are all stable—bans aren’t as severe as rumored. Users considering Claude subscription can reference this suggestion today when choosing payment channels.

⌘ Top Open Source Projects

anthropics/claude-plugins-official: Official Plugin Directory

Anthropic-managed high-quality plugin collection. anthropics/claude-plugins-official core code now public , gained 83 stars today, total 36936 stars. This is Anthropic’s official Claude Code plugin directory. Developers looking to extend Claude capabilities can find trusted plugins from this repo today.

anthropics/skills: Agent Skills Library

Public code library makes agent skills reusable. anthropics/skills core code now public , gained 189 stars today, total 178318 stars. This is Anthropic’s maintained public code library for agent skills. Developers building agent applications can check today for directly callable skill modules.

androoAGI/starnet: Real-Time Pixel Art Workstation

Local-first desktop agent platform. androoAGI/starnet core code now public , gained 93 stars today, total 475 stars. This is a real-time pixel art workstation where real AI agents perform real work. Comes with built-in keys, lets you observe your team operating in real-time. Developers who like pixel art or want to try local agent collaboration can deploy and experience it today.

◉ Social Media Picks

Opus 5.5 Intelligence Comparison Video Shows Significant Leap

Enlightenment-level intelligence jump. XiangYangQiaoMu retweets that ego’s Opus 5.5 comparison promo video shows significant intelligence boost . Claude Opus 5.5’s performance in the video is described as “enlightened”, the comparison quite brutal. Developers wanting to see new model capability changes firsthand can watch this demo today.

DeepSeek Harness Official Client Pairs with Open-Source Browser Plugin

Best open-source browser control plugin. XiangYangQiaoMu introduces in a tweet that after DeepSeek Harness released official GUI client, it can pair with OpenCLI-MCP . This is a brand new MCP 2.0 architecture browser control plugin with experience matching codex plugin but faster. Developers needing to add browser control to Harness can try this combo today.

😄 AI Fun Facts

AI Finally Learns to Multitask

You ask AI to look something up, then immediately remember you need to add something, but it’s still there seriously “searching”—this awkwardness might soon be over. Ant and Tsinghua just open-sourced Realtime-Venus (9B model) which uses a Harness framework to achieve asynchronous parallel processing of dialogue and tasks —you talk your talk, it does its thing, no interference. Finally don’t have to hold your breath waiting for it to finish processing the last sentence.