10-16-daily
AI 日报中心
爱窝啦 AI 日报 AI Daily 2026/10/16: Meta Open-Sources 816B-Parameter Model That Runs Locally on Two Macs
每日 AI 行业动态、产品更新、前沿研究、开源项目与社媒信号
Today’s Summary
Meta open-sources an 816B-parameter model that runs locally on two Macs; DeepSeek Harness gains a context-usage visualization plugin.
Desktop inference, research logging, and an agent skills framework are all solving the same problem: pulling AI capabilities out of demos and into real workflows.
Start today by trying Muse Spark Open and the xerj indexing tool, then ask yourself whether your own context window is overdue for visual management.
` `` `
## **🔥 Today's Top 10 Highlights**
### 1. Meta Open-Sources 816B-Parameter Model That Runs Locally on Two Macs
**You can now run a massive model right on your desktop.** Meta [announced on the official blog that Muse Spark Open is now open source](https://research.meta.ai/blog/muse-spark-open), with **816B** total parameters and 42B active parameters. It natively supports text, image, video, and audio 🎬, with a context window of **1 million tokens**. Deeply optimized for Apple MLX, it can run locally on two M5 Ultra Mac Studios. Designed specifically for coding and multi-step reasoning, it suits long-horizon tasks and complex software engineering.
### 2. AIPOCH and 76 Institutions Jointly Open-Source a Full Research-Process Logging Solution
**AI experiment handoffs no longer rely on verbal summaries.** The **AIPOCH** team [introduced the Open-Science open-source solution on Quantum Bit](http://mp.weixin.qq.com/s?__biz=MzIzNjc1NzUzMw==&mid=2247933984&idx=4&sn=ece2f3d1fe454c8594b75fa22f31f234), in collaboration with **76 research institutions** 🔬. It provides complete records of data sources, code versions, and analysis steps, addressing the collaborative pain point of "which dataset produced this figure and how many times was the code revised." Suited for multi-person iterative AI labs.
### 3. DeepSeek Harness Plugin Makes Context Usage Visible
**You can finally see where your context is going.** The [dsh-context plugin was recommended on Juejin](https://juejin.cn/post/7696710972372009010), adding visibility capabilities 📊 to DeepSeek Harness. It breaks down context by composition and step, showing specific content, **token usage**, and tool calls 🔍. After a multi-step task run, you can **pinpoint which part is consuming the most**. Developers can use this to trim tool configurations or shorten rule sets.
### 4. Anthropic's Officially Maintained Plugin Directory Tops the Daily Rankings
**Claude Code now has an official plugin catalog.** The [Anthropic officially maintained plugin directory topped the daily rankings](https://github.com/anthropics/claude-plugins-official), gaining **307 Stars** ⭐ on the day, with a total of 39,326 Stars. It is a high-quality plugin directory 📦 managed by **Anthropic** directly. Teams using Claude Code can filter for reviewed extensions as needed.
### 5. Gemini 4 Argon Autonomously Builds a 3D Scene in Minutes
**Generative AI is starting to build worlds by hand.** Google AI [shared a hands-on demo of Gemini 4 Argon](https://x.com/GoogleAI/status/2110744759111213217), where the model acts as a virtual builder 🏯 inside Antigravity. It autonomously created a multi-tiered pagoda in **minutes** and **supports full exploration**. Rather than generating a static image, it produces an interactive 3D space. Suited for rapid prototyping in games, architectural previsualization, and virtual-scene validation.
<video controls preload="metadata" playsinline style="max-width:100%; height:auto;" src="https://video.twimg.com/amplify_video/2110743781985198080/vid/avc1/1920x1080/tBWjGxafV6qLt5F0.mp4?tag=29"></video>
### 6. User Clarifies That a Developer Joining Anthropic Is a False Rumor
**Verify the original post before you retweet.** [A Jike user pointed out](https://m.okjike.com/originalPosts/6ad0c5afbd0563695b974e57) that multiple groups were circulating the claim that "a certain developer has gone to Silicon Valley to join **Anthropic**" 📱. However, the original post contains no such statement. This is a reminder that personnel-change news must be checked against the original source to confirm whether it actually mentions a company, role, or location.
### 7. Manus Recruiting AI Video Production Interns in Shanghai
**AI video roles are still expanding.** Chenran [posted an internship opening on Jike](https://m.okjike.com/originalPosts/6ad08bdfb9083720add94415) 💼, based in Shanghai. The role explores the ceiling of AI video production, with above-average intern pay. Interested candidates can email chenran@**manus**.ai or reach out directly on Jike.
### 8. Qianwen Office's Real-Time Second Counter Draws User Complaints
**Showing the elapsed time isn't necessarily a good idea.** Yunfan [complained on Jike that Qianwen Office](https://m.okjike.com/originalPosts/6ad0a356e96dfbdd3025604f) displays how many seconds an Agent has been running in real time ⏱️, arguing that this transfers engineering anxiety onto the user. Other tools simply show "**Generating…**"; watching the seconds tick by makes users focus on speed rather than results. Product managers designing progress indicators may find this **genuine user feedback** worth referencing.
<video controls preload="metadata" playsinline style="max-width:100%; height:auto;" src="https://videocdnv2.ruguoapp.com/FvdRTab_cfmS-c0xWr9NBIav5lzG.mp4?sign=6c98a31cf90527924b67615e67cb6b52&t=6ad1a325"></video>
### 9. User Has AI Design a Custom Cover for a Long-Idle Kindle
**AI design is reaching niche personal use cases.** Linmi [shared on Jike](https://m.okjike.com/originalPosts/6ad059ab0038b7cd5e9da00d) an AI-generated custom **Kindle** cover 🎨. Although the device had been sitting unused for six or seven years, it still received a custom design. This illustrates generative AI's practical application in personalizing everyday objects.
### 10. Twitter User Shares "AI Best Friend" Short Video
**AI interaction short videos are appearing on social platforms.** Gorden Sun [posted a video on Twitter](https://x.com/Gorden_Sun/status/2110758191256715504) titled "**AI best friend, so adorable**" 📹, showcasing a form of AI interaction. See the original post for the full content.
<video controls preload="metadata" playsinline style="max-width:100%; height:auto;" src="https://video.twimg.com/amplify_video/2110758004501192705/vid/avc1/1920x1080/hEGgLLmtVGed48ZQ.mp4?tag=29"></video>
---
## **⚡ Product & Feature Updates**
### AI Exploration Guide Compiles 25 Illustration Prompt Sets and Releases a Skill
**A curated illustration prompt collection is now available.** AI Exploration Guide [compiled 25 hand-picked illustration-style prompt sets in its Telegram channel](https://t.me/aigc1024/25890) 🎨. A Muyang Illustration Skill was also created and published on GitHub. The illustrations from this collection have previously been viewed by over a million people. Creators who need to generate illustration-style images in bulk can use them directly.
---
## **◎ Industry Changes & Personal Impact**
### Cancer Patient Shares Experience with the U.S. Healthcare System and AI's Limitations
**AI reassurance cannot substitute for diagnosis.** JJ Maxwell [shared lessons learned after being diagnosed with cancer](https://jjmaxwell4.com/writing/things-to-know-when-you-get-cancer/), having been diagnosed with oral squamous cell carcinoma in May 2026. His wife's insistence on calling again moved his appointment three weeks earlier. He proactively requested genetic testing but was bounced between specialists, causing a delay of several months. He used AI to organize information and understand his options, but explicitly acknowledged that prior to his diagnosis the model had repeatedly reassured him it "didn't sound like cancer." A reminder to both developers and users: AI can help you understand medical information, but it cannot serve as a diagnostic tool.
---
## **⌘ Top Open-Source Projects**
### obra/superpowers: Agent Skills Framework and Development Methodology
**A structured capability system for your agents.** [obra/superpowers published its skills framework and methodology](https://github.com/obra/superpowers), gaining **582 Stars** ⭐ on the day, with a total of 300,075 Stars. It helps developers formalize agent capability definitions and invocation workflows 📋. Suited for teams building multi-skill agent systems.
### lightseekorg/tokenspeed: Lightspeed LLM Inference Engine
**A new option for inference speed optimization.** [lightseekorg/tokenspeed focuses on inference acceleration](https://github.com/lightseekorg/tokenspeed) ⚡, gaining **30 Stars** on the day, with a total of 2,548 Stars. Worth trying for real-time application developers who are sensitive to inference latency 🚀.
### xerj-org/xerj: Automated Indexing for AI Search Data
**Let your agents stop depending on grep to find data.** [xerj-org/xerj automatically indexes code, documents, logs, and PDFs](https://github.com/xerj-org/xerj) 🔍, gaining **715 Stars** on the day, with a total of 5,876 Stars. Supports search, RAG, security auditing, and agent memory 💾. A single command completes the index, using **40× fewer tokens** than grep. Compatible with Elasticsearch, so existing clients can connect directly. Suited for development teams that frequently search large volumes of unstructured data.
---
## **◉ Social Media Picks**
### Kling 4.0 Officially Released with Magic Card Demo
**Kling's video model has iterated to version 4.0.** Baoyu [retweeted the Kling 4.0 release announcement on Twitter](https://x.com/dotey/status/2110783975245426929) 🎬. The official showcase is a magic card demo video highlighting the new version's capabilities 🃏. Kling 4.0 is now fully available for creators to use.
<video controls preload="metadata" playsinline style="max-width:100%; height:auto;" src="https://video.twimg.com/amplify_video/2110722073765285888/vid/avc1/3840x2160/0fUeia1cVxpB-P1F.mp4?tag=29"></video>
### Kimi Code Theme Color Name Sparks Discussion
**Even product naming can become a talking point.** Xiangyangqiaomu [mentioned on Twitter](https://x.com/vista8/status/2110781330929881337) that one of Kimi Code's theme colors is called "Bright Side of the Moon" 🌙, sparking retweets and discussion. It reflects the care developers put into naming minor details, and shows that users do notice these non-functional design choices.
## **😄 AI Fun Facts**
### Two Macs Can Now Run an 800B Model
Meta's newly released **Muse Spark Open** has **816B** total parameters yet can [run locally on two M5 Ultra Mac Studios](https://x.com/op7418/status/2110786960080155103). This open-source model, deeply optimized for Apple's **MLX framework**, has only 42B active parameters, supports a **1M-token context**, and accepts full multimodal input. Meta calls it "the most powerful open-source model ever built in the United States," and its benchmark scores are on par with K3 and GLM-5.3 — though you will need to scrape together 1 TB of RAM first, so your wallet may need to go on a diet.
## **❓ Related Questions**
### What is the claude-plugins-official project?
This is the officially maintained Claude Code plugin directory managed by Anthropic. It catalogs high-quality, reviewed plugins for extending the functionality of the Claude Code CLI. Unlike community plugin marketplaces, this repository contains only officially certified and managed plugins, ensuring quality and compatibility standards. You can find the list of available plugins and their installation instructions in the [claude-plugins-official](https://github.com/anthropics/claude-plugins-official) repository.Last updated on