[KongchangAI]
· 3 min read· 1,794 words

Gemini 4 Hydrogen Opens to All, GLM-Nano Scores Leak, a Wild Day in AI

Gemini 4 Hydrogen Opens to All, GLM-Nano Scores Leak, a Wild Day in AI

Google Search fully integrates Gemini 4 Hydrogen, NVIDIA open letter story twists, Qwen4 marketing draws mockery.

This AI news roundup covers major developments. Google Search gets its biggest model update ever, fully integrating Gemini 4 Hydrogen for global users with complex question understanding, typo recognition, visual answers, and parallel sub-topic queries—search is shifting from links to structured answers. The NVIDIA employee open letter was confirmed to reuse an old OpenAI template, but an internal poll shows over 97% want Jensen Huang to stay as CEO, with over 1,600 replies. Alibaba's Qwen4 official version is imminent but still marketed via preview, drawing complaints; the Codex reset card has arrived. Cloud and Microsoft cloud services also suffered consecutive outages, DeepSeek Harness updated to V0.22 RC3 with a new drawing import plugin, and GLM-Nano benchmark screenshots circulated from a single source. Tencent Hunyuan HiRealTime, World Labs ALICE, Devin Teams, and Perplexity's open-source models also saw new progress.

Google Search gets its biggest model update ever as Gemini 4 Hydrogen opens to all users worldwide; the NVIDIA employee open-letter story takes a twist, though an internal poll shows over 97% support for Jensen Huang; GLM-Nano benchmark scores appear to leak, while Qwen4's official release is imminent but still shrouded in teasing. This episode of AI news is packed with high information density—let's break it down one by one.

Google Search Fully Integrates Gemini 4 Hydrogen, Rebuilding the Search Experience

Google announced that Search is receiving its largest-ever model update, fully integrating Gemini 4 Hydrogen and rolling it out to all users globally starting today. Key updates include: better understanding of multi-part complex questions, support for typo recognition, and coverage across all AI Mode languages. The answer presentation is also shifting toward a visual layout, supporting images and comparison tables, with query capabilities upgraded to parallel sub-topic queries that can simultaneously link to more relevant web pages. This means search results are evolving from link lists toward structured answers.

Interestingly, Google is pushing Hydrogen to lower-tier Pro users to ensure the vast majority can use it. Currently, most users don't have access; Pixel and student subscriptions should be available. Separately, a Bilibili creator mentioned in a video that Anti-Gravity has now launched Orchestrator mode, with the Ultra plan powered by Carbon and the Pro plan by Argon. This mode lets a large and a small model collaborate dynamically on the same task to reduce cost and improve efficiency. It's worth distinguishing that Anti-Gravity's web and app versions have always been available—this rollout restriction mainly affects specific subscription tiers.

Over 1,600 internal NVIDIA employee replies

AI Mode is the search format Google Search introduced at the 2025 I/O conference. After users enter a question in the search box, they enter a standalone interface centered on conversational AI, where the Gemini model directly generates a comprehensive answer rather than returning a traditional list of blue links. It coexists with traditional search, with its entry point at the top of the results page. AI Mode supports multi-turn follow-ups, multimodal input (images, voice), and parallel sub-topic queries, relying on query fan-out technology: breaking a single complex question into multiple sub-queries retrieved in parallel, then aggregating them into a structured answer. The integration of Gemini 4 Hydrogen means AI Mode is fully upgraded in language coverage, typo tolerance, and visual presentation, and it also marks Google Search's product focus shifting from "retrieving links" to "generating answers."

The Jensen Huang "Reinstatement" Drama Takes a Twist: Open Letter Uses Old Template, but Internal Support Is Stunning

Reports circulated that hundreds of NVIDIA employees signed an open letter angrily criticizing the board as incompetent, demanding the board resign and immediately reinstate Jensen Huang as CEO, or else they would join Elon Musk's SpaceX TerraFab project. However, verification revealed that this so-called open letter actually used the template of the public letter OpenAI employees signed in November 2023 when Sam Altman was ousted by the board and they demanded his reinstatement as CEO.

But the story isn't entirely baseless. According to Phoenix Technology, in a poll initiated by employees on NVIDIA's internal network, over 97% of voters want Jensen Huang to continue as CEO. Among internal suggestions, a post by NVIDIA employees received over 1,600 replies, with most employees supporting Huang's return. Musk also reposted an A16Z post praising Huang: "You will be remembered forever, because some things are eternal."

In the background, Jensen Huang is seen as a key figure in relaxing chip export restrictions to China, and NVIDIA's heavy investment in OpenAI has led OpenAI to consistently support relaxed chip exports to China. As Democrats seized power in the midterm elections, Huang was forced out, and OpenAI is joining Anthropic in strongly supporting chip export controls and pushing to pass legislation by year-end. Sam Altman and Dario have finally "joined hands." Additionally, rumors suggest DeepSeekAd V5 is being trained entirely on Huawei chips and may launch in Q1 next year, with Huawei Ascend 960DT expected around the same time.

Codex Reset Card Arrives, Qwen4 Official Release "Coming Soon" Draws Mockery

The Codex reset card arrived after 6 AM today; users should check for it. Teebo posted asking how to improve Codex, specifically noting "please don't mention new models." A subsequent reply from Teebo containing the word "China" caused confusion, later revealed to be old and new Teebo playing a meme—the new Teebo's surname ends in Brosino.

Old and new Teebo meme screenshot

Alibaba Cloud officially announced that Qwen4 is coming soon. The 2026 Alibaba Cloud Hong Kong Summit begins on November 17—could Qwen4 Max official version be released there? But Qwen4 Max Preview runs until November 25; if the official version launches, the preview becomes awkward. The video creator bluntly called Alibaba's marketing "absurd": after long hype, only a preview was released; earlier the answer was "coming next week," now it's "coming soon," and across a large and small model plus four series, not a single official version exists.

Qwen4 marketing mockery screenshot

Preview vs. General Availability (GA) are two common stages in large model releases. Preview versions are typically opened to developers or early users, may have incomplete features and performance, temporary API pricing or quotas, and no SLA commitment; GA means model weights, interfaces, pricing, and service levels are finalized and ready for production. Vendors often use Preview to build hype and collect feedback, but if the preview period drags on while GA is delayed, it easily triggers user frustration over "marketing first, delivery lagging." In this article, the Qwen4 Max Preview expiration date is close to the Hong Kong Summit date, which is the key clue for judging the GA release timeline: if GA launches at the summit, the preview's existence lasts only a few days, making it awkward; if delayed, the marketing promise falls through.

Cloud Service Outages and a KYC Interlude

A user reported that using a Chinese passport plus a Japanese visa passed Cloud Code KYC instantly; users with a Japanese visa can try, but if worried, find someone to do it for you. Cloud Code's web and API platform services experienced identity verification failures, and Cloud servers crashed extensively, lasting nearly two hours from 6 PM. Microsoft also went down, lasting 17 hours. The video creator joked: one is IPO-ing, the other is doing mass layoffs—can't you pay a little attention?

DSH Update and Model Leaderboard Mystery

DeepSeek Harness has been updated to V0.22 RC3, adding an experimental DeepSeek drawing import plugin that can import historical drawings from the DeepSeek app or web version into the workspace to continue chatting; it also fixed issues where in-app updates on the Windows desktop version might fail and the app might not reopen automatically after updating.

DeepSeek drawing import plugin

A screenshot of an AI model leaderboard is circulating online, showing a mysterious GLM Nano series with three versions: Pro, MAD, and Small Pro, reportedly surpassing Opus 5.5, MAD, and Astra. If true, the GLM 5.5 series would be a leap forward. However, the leaderboard comes from a single source and should be viewed with caution.

3D Gaussian Splatting is a 3D scene reconstruction and rendering technique proposed in 2023. It represents a scene with millions of 3D Gaussian ellipsoids carrying position, color, opacity, and covariance parameters, then renders images in real time via differentiable rasterization. Compared with NeRF, it has order-of-magnitude advantages in training speed and real-time rendering frame rates, and has become a mainstream solution for quickly generating interactive 3D scenes from multi-view photos. World Labs' ALICE lets users generate keyframes, video clips, or 3D Gaussian splats from text ideas—essentially combining generative models with a 3D reconstruction pipeline: first generate multi-view consistent images or videos, then reconstruct them into a navigable 3D representation. Such outputs can be used directly in games, film previsualization, and spatial intelligence training.

Other Updates: Hunyuan, World Labs, Devin, and Perplexity

Tencent Hunyuan released a new real-time voice model, HiRealTime, which in addition to foundational capabilities in voice understanding, instruction execution, and anthropomorphic expression, adds Agentic tool calling and singing abilities. It is already live in Yuanbao voice calls and Tencent customer service, and will soon be offered as an API service on Tencent Cloud; it leads in comprehensive objective performance on multiple public benchmarks. World Labs announced that ALICE is now open to everyone, providing an app and API endpoints where users can submit ideas for building 3D worlds and receive keyframes, video clips, or 3D Gaussian splats. Cognition upgraded the Devin Teams plan: team members can use the new Devin Desktop without extra seats, SWE2 free time in Devin is now permanent, and Teams also supports two free Cloud parallel completions—considered one of the most cost-effective AI subscriptions available. Perplexity open-sourced decision models Decider V1.3 and Flash V1.3, fully available on Hugging Face and free for one week on Open Router.

Agentic tool calling means that a voice or dialogue model, after understanding user intent, no longer just returns a text answer but actively calls external tools (such as calendars, search, payments, booking APIs) to complete multi-step tasks. It requires the model to have function calling capabilities, task planning and state tracking, and to handle low latency and interruptions in voice scenarios. The Agentic tool calling added to Tencent Hunyuan HiRealTime means Yuanbao voice calls can upgrade from a "Q&A assistant" to an "assistant that gets things done"; combined with singing ability, it shows expansion in prosody, emotion, and timbre control. Leading on public benchmarks usually refers to composite scores on standardized tests such as speech recognition, semantic understanding, and instruction following, but real-world experience also depends on end-to-end latency and the richness of the tool ecosystem.

One-line Summary

Google Search fully integrates Gemini 4 Hydrogen, the NVIDIA open letter story twists, and Qwen4 marketing draws mockery.

Paragraph Summary

This episode of AI news covers several important developments. Google Search gets its biggest model update ever, fully integrating Gemini 4 Hydrogen and rolling it out to global users, with support for complex question understanding, typo recognition, visual answers, and parallel sub-topic queries—search is shifting from link lists to structured answers. The NVIDIA employee open letter was confirmed to reuse an old OpenAI template, but an internal poll shows over 97% of employees want Jensen Huang to remain CEO, with over 1,600 replies. Alibaba's Qwen4 official version is imminent yet still marketed via preview, drawing user complaints; the Codex reset card has arrived. In addition, Cloud and Microsoft cloud services suffered consecutive outages, DeepSeek Harness updated to V0.22 RC3 with a new drawing import plugin, and GLM-Nano benchmark screenshots circulated but from a single source. Tencent Hunyuan HiRealTime, World Labs ALICE, Devin Teams, and Perplexity's open-source models also saw new progress.

#Gemini 4 Hydrogen#Google Search AI Mode#NVIDIA Jensen Huang#Qwen4#GLM-Nano#DeepSeek Harness#Tencent Hunyuan HiRealTime#World Labs ALICE
Share:

Related articles