2026-11-17 12:10
Introduction: The 594B Model and Mobile Agent, Open Source Accelerates Workflow Takeover. This Week in AI Is Absolutely Explosive.
This week in the AI ecosystem has been absolutely explosive.
Wenxuan Huang suddenly stepping down as CEO sent shockwaves through the compute market; meanwhile, model giants showed no signs of slowing down—Google unveiled Gemini 4 Hydrogen, launching a new round of low-cost blitz, followed closely by Anthropic updating its Claude 5.2 series.
The real developer community is proving highly pragmatic—once-exclusive features behind closed-source APIs such as native 4K image generation and ultra-large-scale MoE programming models are now open-sourcing weights just days after release. Agents are fully exiting chat interfaces, leveraging MCP and increasingly mature Skill ecosystems to take control of mobile screens, integrate into team collaboration workflows, and even package top-tier institutional thinking methodologies into reusable skill kits.
In this episode of Digging GitHub Weekly, we’ve curated 14 of the most representative open-source projects from the past week’s rising stars—from the newly released 594B programming model and 4K image model, to MCP tools capable of operating your smartphone, Jane Street’s open-sourced problem-solving Skill, and even a browser-based vocal synthesizer built on pure physical modeling.
01. Open-Source Models
MiniMax released H3-Image on November 11, unifying text-to-image generation, reference-image synthesis, and image editing within a single model. The model employs a 6-step sampling process, directly generating images up to 4096 × 4096 pixels with support for multiple aspect ratios and up to 16 reference images.
Multi-image input enables composition of characters, products, and scene elements, while editing capabilities allow modifications based on textual instructions. Ideal for poster design, product imagery, character and environment creation.

Model weights are available for download; local deployment requires substantial GPU memory and supporting components.

Zyphra’s ZONOS3-Mini is a multilingual speech generation model based on a sparse MoE backbone, with approximately 800 million total parameters, activating around 100 million parameters per token, supporting 43 languages including Mandarin.
The model can generate voice samples with similar timbre to reference recordings, enabling control over speech rate, emotion, and vocal expression—ideal for video narration, character dubbing, and voice interaction development.
https://huggingface.co/Zyphra/ZONOS3-Mini
Thinking Machines Lab released Inkling 2 Small on November 5 and opened its weights on November 12. The model uses a sparse MoE architecture with 594B total parameters, activating 17B parameters per token, supporting text and image inputs with text output, designed for programming, tool invocation, information retrieval, and multimodal application development.

The model provides NVFP4 and MXFP8 weight variants, compatible with inference frameworks like SGLang and vLLM, licensed under Apache 2.0. The NVFP4 variant requires at least 500 GB of total GPU memory, primarily targeting research and development teams with multi-GPU servers.
02. AI Applicationshttps://huggingface.co/thinkingmachines/Inkling-2-Small
Pinky: Connecting Mobile Devices with AI Agents

Pinky is an open-source mobile operation app that enables Agents to perform screen reading, tapping, and input via MCP, allowing tools like Claude Code, Codex, and Cursor to control real iPhones or Android devices.

The project consolidates screen mirroring, operation logs, and human approval in one interface, supporting real-time viewing of execution processes, playback of recorded sessions, and immediate takeover. Actions like payments default to confirmation requests, while message sending and deletion can be configured to prompt first—ideal for experimenting with mobile automation and mobile app testing.
Pinky offers macOS, Windows, and Linux clients; mobile connectivity and authorization require system-specific setup, and the project remains in pre-alpha stage.
Project URL: https://github.com/DIYgod/pinky
agentes.chat is a self-hostable team collaboration platform integrating channels, private chats, files, tasks, and AI Agents. Each Agent can be configured with roles, models, budgets, and owners, and connected to external tools via MCP.

Teams can assign tasks, track progress, schedule recurring work, and confirm sensitive operations via approval cards. Ideal for sharing AI workflows across teams.
03. Creative ToolsProject URL: https://github.com/agentes-chat/agentes-chat
CHIMERA generates human voices by simulating airflow, vocal cords, vocal tract, and lip mechanics. The Rust audio engine runs via WebAssembly in the browser. No voice samples, no AI voice models—directly simulates lungs, vocal cords, tongue, lips, and nasal cavity to let computers produce natural speech!
Users can drag the tongue, lips, and jaw to observe real-time vocal changes, input text or lyrics for reading, singing, or rhythmic performance. Real-time visualization displays airflow, vocal cord vibration, spectrum, and formant movements during phonation.
Project URL: https://github.com/jamesondh/chimera-vocal-synth
GIF Factory organizes creation instructions, style references, and scripts into Agent Skills, covering meme content, animated posters, signature animations, and dynamic storybooks.
Provide character art, script drafts, dialogue lines, and motion requirements—Agents follow defined workflows to generate GIFs or videos.

Ideal for turning existing characters, pet photos, and brand assets into dynamic content, requiring an Agent environment that supports image generation and local scripting.
Project URL: https://github.com/WoYun520/gif-factory
Claude Animated Shorts is a set of animation skills tailored for Claude Code, placing 2D characters into 3D scenes and orchestrating shots, actions, music, and sound effects using Three.js and Remotion.
Begin with a story theme—let AI break down plot points and design shots—then use code to animate character movements and scene transitions.
Best suited for scientific education, cartoon storytelling, and animated short films.
04. Productivity ToolsProject URL: https://github.com/mromidev/claude-animated-shorts
Yogurt is Frigade’s open-source AI meeting note-taking application, capable of capturing microphone and system audio, transcribing via Whisper and other speech models, then combining manual notes and local LLMs to generate summaries. All processing—recording, transcription, and summarization—occurs locally, eliminating the need for recording bots during meetings.

Supports cross-meeting search, Q&A based on records with citations, and export of notes and transcripts as Markdown. Official installers are available for Apple Silicon Mac, Windows, and Linux, requiring at least 8 GB RAM; users must download local models on first use.
Project URL: https://github.com/FrigadeHQ/yogurt
An open-sourced learning Skill from Jane Street, a top-tier Wall Street quantitative firm. It begins with foundational diagnostics, breaks down knowledge units, advances learning one question at a time, tracks mastery levels across topics, and incorporates spaced repetition and final assessments.

Starts with 3–5 diagnostic questions. Only one question at a time is presented. If incorrect, it doesn’t reveal the answer immediately but asks a narrower follow-up to help identify contradictions—only after two or three mistakes does it provide direct explanation.
Can be used directly in Claude Code and other Skill-enabled tools; paste content into any chatbot for basic functionality.
Project URL: https://github.com/janestreet/agent_skills
HolyArrs is an open-source self-hosted media library management platform that centralizes movie/series requests, version filtering, subtitles, download tasks, and cataloging—all within a single web interface.

Users can define rules by resolution, file size, and language, integrate with existing downloaders and media servers like Jellyfin and Plex—ideal for managing home media libraries on NAS or server environments.
Project URL: https://github.com/HolyArrs/HolyArrs
05. Development ToolsKiri is a lightweight container runtime written in Rust, compatible with Docker Engine API, allowing reuse of Docker commands and Compose configurations, and providing a graphical interface for managing containers, images, data volumes, and networks.

Supports macOS, Windows (via WSL 2), and Linux—ideal for individual developers and small teams managing local development environments using Docker commands and Compose.
Project URL: https://github.com/libracoder/kiri
gputaskman is a terminal monitoring tool for local GPU workloads, displaying process-level GPU memory usage and activity, further labeling running frameworks, model files, and estimating weight and cache consumption to assist in diagnosing local model GPU memory usage.

Project URL: https://github.com/shanselman/gputaskman
DeepSeek’s open-source chat-demo provides a chat interface powered by API, exposing functionalities like search, file reading, link parsing, image recognition, speech-to-text, and text-to-speech to developers via API.

With the source code, integrate with DeepSeek’s API to customize the UI, add features, and build your own AI application.
API Key, settings, and chat history are stored locally in the browser. When sending messages, attachments, or using voice functions, corresponding content is submitted to the configured API service. Ideal for self-hosting chat interfaces or further developing chat applications.
Project URL: https://github.com/deepseek-ai/chat-demo
Original Article: Issue 3 of Digging GitHub Weekly: Gemini 4 Just Got New Updates! MiniMax Opens 4K Image Model
Source: AI Tools Kit
Disclaimer: Contains third-party opinions, does not constitute financial advice
Step Code – The Open-Source Terminal Programming Agent by Jieyue Star
09-24
Hy4 Preview Lightweight Version – Tencent Hunyuan's Open-Source Hy4 Quantized Compression Model
09-02
Claude Fable 5.1 – Anthropic's Latest Flagship Large Model
09-02
Open-source Retool is here, drag-and-drop backend with 60+ components
08-19
Claude Opus 5 is Here: Half the Price Approaching Fable 5
07-25
Tencent Open-Sources 260 Real-World Tasks to Test AI Workforce Capabilities
07-25
Meituan Open-Sources AI Exam Platform: 813 Minecraft Tasks
07-25






