2026-11-15 15:40

Introduction: The integration of million-Token models and local Agents is streamlining workflows. Digging GitHub Weekly is a weekly curated showcase of open-source projects, primarily highlighting GitHub repositories with rapid Star growth over the past week, covering AI applications, productivity software, and development tools.

Digging GitHub Weekly is a weekly curated showcase of open-source projects, primarily highlighting GitHub repositories with rapid Star growth over the past week, covering AI applications, productivity software, and development tools.

01. Kimi K3.1 Pro: Weights Released This Week

The full weights for Kimi K3.1 Pro are now available in the official model repository. The model supports both text and image inputs, featuring a context window of approximately 1 million Tokens, enabling tool-assisted processing of code, visual content, and multi-step tasks.

The official demo showcases generating 3D assets from text descriptions, rough sketches, or reference images. For users interested in similar capabilities, initial experimentation can be done via the Kimi product or API; teams planning self-deployment can refer to inference documentation in the model repository.

Kimi K3.1 Pro has a total parameter count of 2.8T, requiring substantial computational and memory resources for self-hosting.

https://huggingface.co/moonshotai/Kimi-K3.1-Pro

02. Ripple-65B Preview: On-Premise Programming Model

Ricursive Intelligence’s Ripple-65B Preview is a dense language model with 65 billion parameters, designed for programming, reasoning, and Agent-driven tasks, supporting up to ~1 million Token context windows.

This release includes instruction-tuned models, base models, and weights in various numerical formats. Teams wishing to run this programming model locally can explore deployment paths via the official repository and integrate it with existing Agent frameworks.

Ripple-65B Preview is licensed under Apache-2.0. Prior to deployment, verify weight format compatibility, available GPU memory, and inference framework support.

https://huggingface.co/ricursiveAI/Ripple-65B-Preview

03. Magpie: Unified Model Switching

Programming assistants like Claude Code and Codex each maintain separate model configurations, requiring manual adjustments when switching models. Magpie consolidates model services and Agent configurations into a single interface, simplifying model selection across different assistants.

Add providers by entering API Keys or linking existing accounts, then select models from the list for each Agent. Once a new session starts, the updated configuration takes effect. For users managing multiple accounts or routing groups, fallback paths can be configured for failed invocations.

Magpie supports macOS, Windows, and Linux, offering access via menu bar, window, and terminal interfaces.

GitHub: yetone/magpie

04. Agent Autoloop: Agent Task Scheduling

Daily note cleanup, task handling upon new Issue detection, can be defined as rules and delegated to Agent Autoloop, which invokes locally logged-in CLI Agents on your machine.

After completing a round of tasks, the Agent exits; for decisions requiring input, queries are sent via Feishu or desktop notifications, resuming once responses are received. The official repository provides templates for scenarios such as morning reports and website monitoring—customization required to match your own paths and rules.

Agent Autoloop requires Node.js, Git, and an accessible command-line Agent.

GitHub: KKKKhazix/agent-autoloop

05. AnyModelBot: WeChat & Feishu AI Assistant

To submit a request via WeChat and have your local assistant continue processing tasks, consider AnyModelBot. The application runs on your personal computer or server, supports WeChat QR code login, and can also integrate with Feishu bots.

Assign different assistants per task, each preserving their default model, workspace, and memory. Assistants can read/write files, execute commands, and prompt users for authorization or additional information when needed. The device or server running the task must remain online.

AnyModelBot offers desktop versions for Windows, macOS, and Linux. Models can be directly connected via your own API Key or routed through Modelflare.

GitHub: modelflare-ai/AnyModelBot

06. stack-diagram: One-Sentence Architecture Diagram Generation

When drafting project descriptions or technical documentation requiring clarification of relationships between frontend, backend, database, and deployment services, feed your tech stack description into stack-diagram to generate branded architecture diagrams.

For example, specifying where a web app is deployed, which API is called, or which database stores data enables the Agent to organize nodes and connections accordingly. Results can be saved as self-contained HTML, exported as PNG, or generated as React components. stack-diagram functions as an Agent Skill, requiring integration with compatible tooling that supports skill execution.

GitHub: caru-ini/stack-diagram

07. Cicada: Real-Time Meeting Subtitle Translation

For listening to English meetings or live streams, Cicada (Zhilei) captures system audio and displays bilingual subtitles; after configuring corresponding services, it can simultaneously summarize minutes, extract discussion points, and track action items.

Post-meeting, users can search transcripts, replay specific segments, and export subtitles in Markdown, Word, or SRT formats. For speaking contributions, there's a "press-to-speak" mode with real-time translated audio output.

Cicada offers versions for Mac, Windows, Linux, and Android. Apple Silicon Macs support native speech recognition; other platforms require cloud service configuration.

GitHub: wquguru/cicada

08. Quiet.md: Local Note-Taking

Quiet.md is a local note-taking application supporting headings, to-dos, tables, code blocks, and full-text search, with notes stored as DocxXML files within user-defined folders.

It allows importing Markdown and Feishu DocxXML, exporting to Markdown or PDF. Ideal for organizing work notes, article drafts, and project materials while retaining familiar document structures.

Official installers are provided for Apple Silicon Mac and Windows; the interface supports Simplified Chinese and is currently in Beta. For multi-device usage, users can manually pair with file synchronization tools.

GitHub: s1dashu/quiet-md

09. sand3d: Python Script-Based 3D Modeling

For creating a bracket with holes, later needing dimension and hole spacing adjustments, let an Agent generate a parametric Python script using sand3d, then modify parameters to produce different variants.

sand3d is built on Blender’s bpy, abstracting operations like outlines, solids, holes, chamfers, and assemblies into a unified modeling API. Generated results can be exported in GLB, STL, and other formats, with source scripts remaining editable.

sand3d is a modeling library requiring a compatible Python and bpy environment—ideal for users willing to script or modify 3D objects.

GitHub: Sandflow-ai/sand3d

10. cinecode: Enabling Agents to Generate Videos via Code

Provide a short film concept to Claude Code or Cursor; cinecode guides the Agent to structure a script and shots, generates visuals and audio via code, and renders the final output as MP4.

Begin with a 30-second product animation or educational clip: define aspect ratio, visual style, and audio tone, review static scene frames, then refine shots, text, and pacing based on the draft.

cinecode requires a local environment including Node.js, Python, FFmpeg, and Chrome; rendering speed is influenced by GPU performance.

GitHub: sarpstar/cinecode

Original Link: Issue 2 of Digging GitHub Weekly: Kimi Releases Weights, Real-Time Meeting Translation

Source: AI Tools Hub