Xiaoliu BOT
中EN
Briefs/2026.11.13 Fri
Issue 226

X Platform November 13 AI Brief | Gemini 4 Hydrogen Opens for Daily Tasks, DeepSeek Opens Search and File APIs, Claude Meetings Beta Launches

01 Gemini 4 Hydrogen Opens for Daily Tasks, Antigravity Adds Dual-Model Orchestration

Google began rolling out Gemini 4 Hydrogen on November 12, positioning it as the fast tier for daily coding and agent work. It is now accessible via the Gemini App, Search, AI Studio, and API; paid Pro/Ultra users can also use it in Antigravity. Google claims it completes tasks 2.5 times faster than 3.8 Flash and offers four levels of thinking depth; these performance comparisons are based on benchmarks published by Google itself. Another change for complex tasks is the Antigravity Orchestrator: Carbon plans, Hydrogen executes code modifications and tests, and if obstacles arise, Carbon adjusts the plan. Google reported in Terminal-Bench 4.0 tests that the success rate is close to using Carbon throughout, while using 24% less quota and being on average 19% faster.

02 DeepSeek Opens API Access to Search, File, and Voice Capabilities Behind Its App

DeepSeek officially announced that the search, file, link processing, security, and voice capabilities from its App are now available via API for combined or individual use, including interfaces for OCR, ASR, TTS, etc. The company also provides open-source frontend examples and code that can import DeepSeek conversation history. This allows developers to directly combine existing capabilities or build their own clients without having to rebuild basic interactions from scratch. Extended tools incur separate usage fees, and specific pricing should be verified against the official price list.

03 Claude Meetings Beta Enables Collaborative Document and Whiteboard Creation During Meetings

Anthropic has begun rolling out the Claude Meetings beta to Team and Enterprise users, initially supporting Google Meet. After joining as a participant, Claude can organize documents, draw on whiteboards, and answer questions while listening to the discussion; after the meeting, users can view minutes, action items, transcripts, and generated content on Claude’s meeting page. Anthropic stated that other meeting platforms will be supported in the future. Baoyu’s detailed post covers permissions, external participants, transcription, and other details, noting an inconsistency between the official announcement article and the help center regarding default settings for meetings with external personnel; administrators should verify their organization’s settings themselves.

04 Detailed Analysis Emerges on MiniMax H3-Image’s High-Resolution Multi-Image Fusion Capability

Xiaohu conducted a product and API-level analysis of the recently open-sourced MiniMax H3-Image: the model can accept up to 16 reference images, with different references specified for different objects; output can reach approximately 16 million pixels, with example sizes including 4096×4096 and 5472×3072. Test scenarios cover furnishing an empty room, multi-person outfit changes, game CG, and structural diagrams; the API can also expand brief Chinese prompts into detailed scene descriptions. The author also notes that the local model requires downloading about 135GB, and the license has restrictions for users in the US, EU, and other regions, as well as for those with annual revenue exceeding $1 million; the prices and examples in the article are compiled by the author, and official terms should be checked before use.

05 Personal Implementations of Fable 5.2 Demonstrate Creating Playable 3D Games from Prompts

Two creators shared their processes of making game prototypes with Fable 5.2: Fu Sheng stated he created a web-based MOBA prototype playable in both landscape and portrait orientations using a single prompt and about 20 hours, with the compressed web package being about 600KB; he explicitly clarified this is only a demo of a single player against NPCs, with networking and commercial game-level details not being the goal. Guizang previously shared a 3D action game, claiming it took an afternoon and the result was packaged into an HTML file of about 30MB. These are single experiences shared by creators and do not represent typical production costs or the model’s stability on complex projects, but they show that natural language-driven prototyping can already deliver playable works.

06 Engineering Path for Running Local Models on Low-Memory Legacy Devices Quantified

Gorden Sun introduced smolmaxxing, based on llama.cpp, which rewrites underlying computations and compresses vocabulary for old TV boxes and iPads, and uses hard disk swapping and preloading of model data to enable local conversational models on devices with only 1GB of available memory. The records state that a third-generation Xiaomi TV Box S runs a 2B Qwen3.5 model at 7.7 tokens per second and a 16B Ling-mini-2.0 model at 5.1 tokens per second; a 2013 iPad Air runs a 1.7B Bonsai model at 24–26 tokens per second. The data comes from experiments and code repositories presented by the author, indicating these are results specific to particular devices and implementations and should not be directly extrapolated to other hardware.

07 Liquid AI Open-Sources Streaming Audio Codec Emphasizing On-Device Low Latency

Gorden Sun provided a technical introduction to Liquid AI’s open-source LiquidCodec: it targets speech models, encoding speech into discrete features that can be restored in real-time, with the design goal of enabling models to generate audio features in parallel, reducing the sequential dependency of traditional encoding. The author reports that on a single Apple M4 Max CPU core, encoding and decoding achieved speeds 361x and 87x real-time audio speed respectively, and states that 4-bit quantization can lower the barrier for on-device execution. The speed and naturalness descriptions here are from the post’s introduction and cannot yet be considered independent cross-device benchmarks.

08 World Labs Opens Access to Atlas World Model for Everyone

World Labs officially announced open access to Atlas, providing an application and API: users can start from an idea for building a 3D world to generate keyframes, video clips, or 3D Gaussian splats. Fei-Fei Li relayed the open access news and expressed anticipation; the raw/report did not list further usage barriers, performance, or quality data, so this section only records its scope of openness and the output forms listed by the official source.

09 Vercel Releases UI SDK for Agent-Generated Interfaces

After Vercel released the UI SDK, Guillermo Rauch described it as a layer of interface behavior components for Agents to call: encapsulating common behaviors like accessibility, performance, and interaction consistency, while not locking generated results into a single high-level framework or fixed component templates. His assessment is that Agents are better suited to combine interfaces from lower-level building blocks according to the scenario; this is a statement on tool design direction, and the tweet did not provide quantitative effect comparisons.

10 A Tech Job Market Report Provides Data on User Research Role Transitions

The 2026 tech job market report shared by Lenny Rachitsky lists stagnant engineering demand, reduced design roles, widespread AI skills, and fewer remote positions as observed conclusions. Regarding user research roles, data from the cited 2,779 U.S. job seekers (who changed jobs between 2024–25) shows: 49% moved to another user research job, 17% transitioned to design, 9% left the field, and about half of those continuing in research went outside the tech industry. This sample and time frame are limited, suitable as a clue to role transitions, and should not be taken as a full industry census.

11 X’s Single-Post Media Limit Observed by Multiple Users to Increase to 20 Items

Multiple users posted on the day stating that X has increased the number of photos/videos that can be uploaded in a single post from 4 items to a maximum of 20, and demonstrated multi-image posting or testing. The report shows user testing and relayed information, with no direct announcement from an official X account observed, so this is treated as a feature change observed on social media, not an independently verified platform statement by this report.

← All briefs →