大爱 AI 日报 AI Daily 2026/10/29: DeepSeek V5 Model Name Appears in Test Code
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
The DeepSeek V5 identifier has surfaced in test code. OpenRouter reveals that a single GB300 can handle one day of requests for the most popular models, with inference costs far below training costs.
A small-parameter OCR model goes open source, a 3D game built in 16 days, and Gemini connects to fitness data — AI is pushing outward from general chat toward vertical tools and personal creation.
Today, prioritize Focus items 2 and 4: confirm whether inference costs affect your deployment decisions, then try MiniCPM-OCR to see if it can replace your current solution.🔥 Today’s Top 10 Focus Items
1. DeepSeek V5 Model Name Appears in Test Code
A new version identifier has been written into the code. A developer found the model ID field using deepseek-v5
in a test file in the DeepSeek Harness repository. The identifier appears as a model name in a mock of the /models endpoint, listed alongside existing models. No official release has been announced, and no launch date has been disclosed 📅.
2. DeepSeek on OpenRouter Consumes One GB300 Per Day
Inference compute demand is lower than expected. Opencode co-founder Dax shared on Jike that the most popular DeepSeek model on the OpenRouter platform requires roughly one GB300 per day of requests (valued at $6–7 million). The top 20 models on the platform together handle roughly 21.2T tokens in throughput, requiring about 4 GB300s at full load 💻. Factoring in load balancing, 6–8 additional units would cover peak demand. The compute consumed by open-source model inference is far less than that of closed-source models and training workloads.
3. A Developer Turned The Little Prince into a 3D Game in 16 Days
AI can now help individuals complete game development. A case study compiled from GitHub browsing shows that someone used AI to adapt The Little Prince into a 3D story-driven game with explorable planets in just 16 days, with the post earning 3.46 million reads 🎮. The article lists 11 viral examples from overseas, covering genres including sci-fi shooters and narrative exploration. Some projects have open-sourced their prompts, making them useful for creators interested in AI-assisted game development.
4. MiniCPM-OCR 1.3B Released as Open Source, Weighing In at Just 1.3B Parameters
A small-parameter OCR model delivers impressive results. The first OCR-dedicated model in the MiniCPM family, MiniCPM-OCR 1.0, has only 1.3B parameters and focuses on recognizing text, tables, formulas, charts, and structured information 📄. Real-world testing shows strong performance despite its 1.3B parameter count, making it well suited for local Agent and knowledge-base use cases. Developers have noted that it is small but capable, with a clear efficiency advantage on vertical tasks.
5. Gemini Connects to Strava to Generate Personalized Training Plans
Fitness data integrated into an AI assistant. Google announced that Gemini can now connect to a Strava account to analyze workout history and turn years of data into a personalized coaching plan 📊. Strava has more than 200 million users across 185 countries. Subscribers can now use Gemini to analyze running, cycling, and other activities 🏃 and receive targeted recommendations. The AI assistant continues to extend from general conversation into vertical use cases.
6. Hongguo Daily Active Users Reach 168 Million, Average Viewing Time Two Hours Per Day
A short-drama app’s watch time surpasses long-form video platforms. A September Nomura Securities report citing QuestMobile data shows that Hongguo’s daily active users reached 168 million in July , doubling year over year. That exceeds the combined total of Tencent Video, iQIYI, Youku, and Mango TV (approximately 155 million) 📱. By total time spent, it has become China’s fourth-largest app. In February of this year, users averaged 125 minutes of daily viewing — longer than Douyin or WeChat. Qin Shuo worries that algorithms are consuming people’s time, while Cheng Lingfeng argues that AI short dramas are only just beginning.
7. Survival LLM Integrates Offline AI with an Emergency Reference Library
Look up first-aid and repair guides even without internet access. Survival LLM is a free, open-source offline assistant that combines a local language model with the Survival LLM integrated offline AI and emergency reference library 📚. Answers cite the actual articles read and highlight relevant passages. The model and reference materials must be downloaded while online in advance; after that, all queries remain on-device 🔒. On devices with 16–31 GB of RAM, it uses roughly 6 GB with the Granite 4.2 8B model, and the full reference library download is about 9 GB. The author tested it with 22 questions about power outages and first aid; the larger model answered 21 correctly on the main points.
8. Ruan Yifeng: AI Has Permanently Changed Programming
An irreversible shift in the development toolchain. In Technology Enthusiast Weekly Issue 416 , Ruan Yifeng concludes that AI has changed not only the programmer’s career but programming itself 💻. Even if a bubble bursts, AI cannot be removed from the development toolchain. A veteran overseas developer laid out the specific ways AI has changed programming in detail. Ruan Yifeng considers the change irreversible and argues that developers need to adapt to a new programming paradigm.
9. Genie 4 Adds Native Sound to Interactive Worlds
Simulated worlds you can not only see but hear. Google DeepMind has added native sound to its real-time interactive worlds , with music, ambient audio, and character dialogue generated dynamically to match each scene 🎮. Users can move the viewpoint with a keyboard or use natural language to change the weather or add characters. The experience is currently available only as a research preview for trusted testers; general users cannot access it yet. For game prototyping or film previs, check the official demo first.
10. Personal System Iteration: Every Input Must Be Paired with an Output
The input-output loop boosts retention. Jike user Ling Youzhi shares the core principle of pairing every piece of input with an output ✍️. Finish a video — write a summary. Finish a film — write a review. Finish a podcast — write notes. Finish a game — write your thoughts. Then share what you have written; it is the lowest-effort way to create leverage. This suits AI enthusiasts and programmers who want to reinforce learning through consistent output.
⌘ Top Open-Source Projects
elayadesign/ai-design-skills: AI Design Skills Collection
A design skills collection drawing significant attention. elayadesign/ai-design-skills gained 2,139 Stars that day, bringing its total to 11,116 🌟. The repository compiles design skills and resources related to AI. It is well suited for designers and product managers looking to understand how AI tools fit into the design workflow. The project is actively maintained, with a lively community.
macro-inc/macro: Unified Team Workspace
Integrates multiple collaboration tools. macro-inc/macro combines email, chat, documents, tasks, agents, calls, and CRM in a single workspace , connecting and sharing AI memory via @ links 🧠. It gained 117 Stars that day, bringing its total to 5,418. Built in Rust, it suits teams looking to reduce the cost of switching between tools. The AI memory feature can associate information across different contexts.
expo/skills: AI Agent Skills for Expo Projects
Designed specifically for Expo developers. expo/skills is a collection of AI agent skills for working with Expo projects and Expo Application Services 🛠️. It gained 90 Stars that day, bringing its total to 3,577. Written in Shell, it suits developers currently building apps with Expo. The skill collection helps AI agents understand project structure and execute common development tasks.
😄 AI Fun
Google Releases Two Flagship Models in Two Weeks — and You Still Can’t Use This One
Sundar Pichai announced on October 28 that Gemini 4 Carbon was live, just two weeks after the previous flagship model, Argon, launched. Carbon topped the Artificial Analysis v5.0 leaderboard , and on Terminal bench 4.0 it left Claude Opus 5.5 and other competitors far behind. There is one small problem: on launch day, you still cannot use it, and the API waitlist keeps growing. Two flagship models in two weeks, users still locked outside — this iteration pace seems built for benchmarks.
❓ Related Questions
What Kind of Training Plan Can Gemini Generate After Connecting to Strava?
Once you link your Strava account, Gemini can read your historical data and generate personalized training recommendations. The integration announced officially on October 28 lets subscribers analyze their activity history directly. If you are looking to open a new account to try Gemini 4 Argon or other paid features, you can check out ready-to-use options like the 大爱·AI账号店 Gemini Pro annual account card .