[中文] | [English] | [Website]

AI-Compass

AI-Compass is an open-source learning and practice ecosystem for the AI landscape. It helps both newcomers and experienced builders navigate major technical directions with curated resources, practical guidance, and a structure that supports long-term study and real project work.

The repository connects foundations with current practice across language and multimodal models, machine learning, computer vision, NLP, recommender systems, reinforcement learning, RAG, agents, GraphRAG, and engineering workflows from training and inference to product delivery.

AI Compass Knowledge Planet coupon

One-line installation through a coding agent

If you use Claude Code, Codex, Cursor, Windsurf, or another coding agent, send it the following prompt:

If AI-Compass is not already available on this machine, clone https://github.com/tingaicompass/AI-Compass.git first. Then follow https://raw.githubusercontent.com/tingaicompass/AI-Compass/main/Install.md to install it as a local knowledge-base skill that both Codex and Claude Code can use. When finished, report the installation path, verification results, and whether the client needs to be restarted.

This prompt instructs the coding agent to clone the repository when needed, run the root installation script, and verify the symbolic links. Afterward, you can ask it directly about AI technologies, models, projects, and weekly updates.

Contents

• Project Positioning
• Submit a Resource
• Weekly Highlights
• Module Guide
  • Blog
  • Code
  • Basic Knowledge
  • Technical Frameworks
  • Application Practice
  • Products and Tools
  • Learning Resources
  • Community and Platforms
  • Enterprise Open Source
• Future Vision
• Star History

All links are directly accessible.

Project Positioning

AI-Compass aims to be a practical, current, and comprehensive AI learning and practice ecosystem. Its nine connected modules offer an end-to-end path from first principles to implementation, product thinking, and ongoing research.

Core Module Architecture

• Blog: structured technical articles, Python fundamentals, algorithm practice, and future LLM, interview, and enterprise-project guides.
• Code: runnable AI examples and demos, including RAG engineering projects that can be inspected, adapted, and discussed with coding agents.
• Basic Knowledge: AI navigation, prompt engineering, model leaderboards, language models, and multimodal models.
• Technical Frameworks: embeddings, training, inference and deployment, evaluation, and RLHF.
• Application Practice: MCP+A2A, RAG workflows, agents, GraphRAG, DeepSearch, NLP2SQL, and popular AI frameworks.
• Products and Tools: AI applications, products, competitions, and project recommendations.
• Learning Resources: courses, articles, academic tools, interview preparation, software, and technical specializations.
• Community and Platforms: learning platforms, technical communities, and ecosystem resources.
• Enterprise Open Source: Datawhale, Huawei, Tencent, Alibaba, and PaddlePaddle resources.

Who It Is For

• AI learners building a dependable map of the field.
• Developers looking for technical references and engineering patterns.
• Product managers researching AI product design and market cases.
• Researchers tracking emerging methods, benchmarks, and open resources.
• Enterprise teams evaluating technical routes and implementation options.
• Job seekers preparing with interview materials and project practice.

Submit a Resource

To submit an AI tool, open-source project, model service, learning resource, technical article, or community platform, open a GitHub Issue with the following information:

Title: Resource submission: <tool or project name>

Tool or project name:
Official website:
Suggested category:
One-sentence summary:
Key features:
- Capability 1
- Capability 2
- Capability 3
Target users / use cases:
Open source / free / paid:
Additional links:
Submitter notes:

Submissions with complete information, stable links, and a clear relationship to AI learning, development, or application practice receive priority. Commercial products are welcome when pricing, free-tier, or trial terms are stated clearly. Inclusion is based on quality, usefulness, and community value rather than paid placement.

Weekly Highlights

When using the repository as a knowledge base, start with latest.md for current questions and INDEX.md for historical lookup. Refer to TEMPLATE.md when creating a new issue.

Past issues

1. NativeMind local assistant, Gen-CLI, and AI tools including Qianyin Manyu voice generation
2. ChatGPT Agent, Kimi 2, Mistral speech models, Grok emotional companionship, Baidu Tizzy, and Youyan digital humans
3. Qwen3 upgrades, ByteDance GR-3 robot, TRAE SOLO, JoyAgent OxyGent, and Z.ai presentation creation
4. Kling Creative Studio, ByteDance Coze Studio and Coze Loop, Tongyi Wanxiang 2.2, GLM-4.5, and Hunyuan 3D world models
5. Fun AI applications: Quin-AI tarot, FateTell AI fortune analysis, and the Love Pet Mailbox companion app
6. Gemini Storybook, open gpt-oss reasoning models, Qwen-Image, RedOne social models, and Xiaomi MiDashengLM
7. Qwen3-Coder, Ollama Desktop, Kimi K2, FLUX.1 Krea, Xiaoxingxu comics, and H+ medical AI
8. Claude Opus 4.1, MiniMax-Speech 2.5, Qwen-Flash, and Google's Jules coding agent
9. RynnVLA, GLM-4.5V, DreamVVT virtual try-on, WeKnora, GitMCP, and NeuralAgent
10. DINOv3, DeepSeek-V3.1, Qwen-Image, Seed-OSS, CombatVLA, and VeOmni
11. Qoder agentic coding, vivo Vision, AIRI, RM-Gallery, and Sim-Agent workflows
12. NVIDIA Jetson Thor, Gemini 2.5 Flash Image, Youtu-agent, Wan2.2-S2V, and SpatialGen
13. PixVerse V5, gpt-realtime, Grok Code Fast, HunyuanVideo, OmniHuman-1.5, and MiniCPM 4.5
14. Nano Banana workflows, AgentScope, Hunyuan-MT-7B, HunyuanWorld-Voyager, and AudioStory
15. Kimi K2, InfinityHuman, 3D AI desktop companions, and Die Die Club virtual companionship
16. ByteDance Seedream 4.0, Qwen3-Max, EmbeddingGemma, OneCAT, and rStar2-Agent
17. CodeBuddy Code, Jimeng 4.0, MiniCPM 4.1, Hunyuan 2.1, Qwen3-ASR, and SpikingBrain
18. Qwen3-Next, Seedream 4.0, FireRedTTS-2, SRPO, and MiniMax Music 1.5
19. IndexTTS2, HuMo, Stand-In, Youtu-GraphRAG, MobileLLM-R1, and PP-OCRv5
20. GPT-5-Codex, Unitree world models, InfiniteTalk, ROMA, and Hunyuan 3D 3.0
21. Nano Bananary, MCP Registry, Tongyi DeepResearch, VoxCPM, and InternVLA-M1
22. TrafficVLM, DeepSeek-Terminus, Qwen3-Omni, Ant Ling, Wan2.2-Animate, and Qianfan-VL
23. Qwen3-Max, Mixboard, Qwen3-VL, Audio2Face, Vidu Q2, and Qwen3-LiveTranslate
24. DeepSeek-V3.2, Sora 2, Imagine v0.9, LONGLIVE, xLLM, and OpenAgents
25. ChatGPT Atlas, Claude Code, Haiku 4.5, Veo 3.1, nanochat, and DeepSeek-OCR
26. Cursor 2.0, Firefly Image 5, Agent HQ, LongCat-Video, and Kimi K2 Thinking
27. Gemini 3, Grok 4.1, GPT-5.1, Qwen, and Lumine-3D open-world agents
28. Nano Banana Pro, Gemini 3, HunyuanVideo 1.5, and Meta SAM 3D
29. Open-AutoGLM, Z-Image, GLM-4.6V, and Kling 2.6 audio-video synchronization
30. GPT-5.2, Qwen3 omni upgrades, Runway world models, and Zhipu's video-generation twins
31. GLM-Claw, EdgeClaw Box, LongCat-Flash-Prover, and FramePraise private beta
32. Veo 3.1 Lite, Qwen3.5-Omni, and DeerFlow 2.0
33. Qwen3.6-Plus, Wan2.7-Video, and Gemma 4
34. Claude Mythos, GLM-5.1, and LifeSim
35. OmniShow, Gemini 3.1 Flash TTS, and Hunyuan 3D World Model 2.0
36. HappyOyster, Qwen3.6-35B-A3B, and Claude Opus 4.7
37. Qwen3.6-Max-Preview, ClawLess, and AgentScope Tuner
38. GPT-5.5, DeepSeek-V4, Spark X2, Tencent offline translation, FlashQLA, and TIPSv2
39. Grok 4.3, Flipbook, OpenLess, OfficeCLI, Career-Ops, and FlashQLA
40. Claude Computer Use best practices, Lumen Flow, AGenUI, General365, InsForge, and agents-cli
41. Gemini Omni Flash, Hy translation, Gemini Spark, Violin, LongCat-Video-Avatar 1.5, and GLM-5.1-highspeed
42. Gamma-World, Claude Opus 4.8, Hermes Desktop, OmniVoice Studio, Bailian CLI, and Qwen-VLA
43. PawBench, Miaoya, Open Code Review, Microsoft Scout, Gemma 4 12B, and Magenta RealTime 2
44. Gemini 3.5 Live Translate, Claude Fable 5, SkillSpector, MiMo Code, HPC-Ops, and EvoQuality
45. GPT-5.6, Qwen-AgentWorld, DSpark, BrowserBC, and Ornith-1.0
46. Claude Fable 5 system prompts, SkillSpector, turbovec, academic research skills, and Qwen-Robot Suite
47. FuckClaude, Claude Science, Hy3, Leap Dimension, and EdgeBench
48. GPT-Live, SayIt, JellyToken, OpenScience, and InternAgentS
49. Kimi K3, Qwen-Audio-3.0-Realtime, Colibri, StaffDeck, and Nemotron 3 Embed
50. MineExplorer and WorkBuddy Bench bring agent evaluation into real tasks; Claude Opus 5 and Yuanji AI connect models to content delivery
51. Orchard, Warp Agent CLI, and WorldClaw accelerate controlled delivery; SmartSub and Wan-Animate-2 expand multimodal production
52. PixelRAG, DeepSeek Harness, Nemotron 3.5 Lightning, and Muse Glimmer shift AI delivery toward efficient, verifiable multimodal agents

Latest Issue

AI Compass Weekly: Qwen3.8-27B and GLM-5.3 Turn Foundation Models into Multimodal and Post-Training Choices, While Grok 4.6 and Gemini 3.7 Flash Raise the Bar for Specialized Workflows

Issue 52 · 2026-08-17 · AI Compass Weekly 21 items · 8 highlights · 3 themes · approximately 3 minutes

Theme

Model capabilities are becoming distinct, verifiable delivery choices.

This week is not simply about larger models; capabilities are separating into deployable choices. Qwen3.8-27B provides native multimodal understanding and long context, while GLM-5.3 uses post-training scaling for coding. Gemini 3.7 Flash targets code and agent workflows, and Grok 4.6 pairs 1.5T parameters with a 500K-token context window. WorkSwarm brings multi-agent collaboration into office and coding spaces, while img2threejs turns a product image into an interactive Three.js asset. Teams should evaluate foundation models, workflows, and output chains separately against context, code, multimodal delivery, and local-execution constraints.

Highlights

Foundation Model Selection Is Splitting into Scenario-Specific Delivery

• Qwen3.8-27B
  • Update: Alibaba's new open model supports native multimodal understanding and long-context processing.
  • Why it matters: Teams that handle text, images, and long documents can evaluate it as a unified foundation-model candidate.
• GLM-5.3
  • Update: Shares its base with GLM-5.2 and uses post-training scaling to improve capability and coding performance.
  • Why it matters: Coding-model evaluations should compare task gains from post-training, not only pretraining scale.
• Gemini 3.7 Flash
  • Update: Targets code development and agent automation workflows, with improved software-engineering and web-development benchmarks.
  • Why it matters: High-frequency development automation can be tested against existing tool-calling, speed, and pricing constraints.
• Grok 4.6
  • Update: Uses a 1.5T-parameter MoE architecture, offers a 500K-token context window, and accepts text and images.
  • Why it matters: Cross-repository and long-horizon tasks need explicit acceptance criteria for context capacity and response speed.

Multi-Agent Systems Are Moving Toward Controlled Collaboration and Local Boundaries

• WorkSwarm
  • Update: Provides office and coding spaces for multi-agent collaboration and reduces token use through context slimming.
  • Why it matters: Organizations can test whether role-based collaboration actually lowers manual orchestration and operating costs.
• wigolo
  • Update: An open, local-first agent network layer that connects coding agents through MCP for multi-engine search.
  • Why it matters: Teams with strict data boundaries can evaluate local retrieval and unified access instead of scattered external-search calls.

Multimodal Creation Is Producing Reusable Production Assets

• MiniMax Music 3.0
  • Update: Generates complete 32kHz stereo songs of up to five minutes from lyrics and structured descriptions.
  • Why it matters: Content teams can establish reusable audio-production workflows instead of treating generation as a one-off ideation tool.
• img2threejs
  • Update: Reconstructs one product image into an interactive, animation-ready Three.js model with TypeScript and JSON output.
  • Why it matters: Commerce, web, and prototyping teams can validate asset editability and front-end integration cost directly.

---

Module Guide

Blog

11.blog / 1.coding programming collection

The blog module develops practical programming ability. Its current core is a structured Python and LeetCode collection: 19 focused Python lessons and 107 high-frequency problem walkthroughs spanning arrays, hashing, two pointers, sliding windows, linked lists, stacks, trees, binary search, backtracking, dynamic programming, graphs, tries, and heaps. Future additions include LLM guides, interview preparation, and enterprise-project explanations.

• LLM guides and interview collection
• Python fundamentals and 107 problem lessons

Code

11.code / Milvus RAG engineering example

The code module contains runnable and adaptable AI engineering examples. The Milvus / Zilliz Cloud RAG demo covers configuration, OpenAI embeddings, Gemini query rewriting, dense-vector and BM25 hybrid retrieval, RRF fusion, Cohere reranking, an LLM answer chain, and end-to-end tests. It is designed both for learning RAG systems and for code-level analysis with Claude Code or Codex.

Basic Knowledge

This module is the repository's entry layer for discovery, prompts, model comparison, language models, and multimodal models. Use it with weeklyHighlights: module pages provide stable knowledge, while weekly issues provide recent additions.

0. AI Navigation Toolset

Use this directory to build a first map of available AI tools before evaluating individual products. It covers navigation sites, developer communities, tool collections, and product discovery resources for writing, coding, image, video, audio, search, and automation workflows.

1.1 Prompt Engineering

This directory moves from asking better questions to reusable, testable, and iterative prompt workflows. It covers role design, context, objectives, constraints, output formats, few-shot examples, reasoning patterns, prompt testing, optimization, safety boundaries, and project rules for AI coding environments.

1.2 LLM Evaluation Leaderboards

Use the leaderboards to compare models by task rather than by a single overall rank. The collection includes general dialogue, coding, mathematics, Chinese-language ability, RAG, multimodal, and agent benchmarks such as LMArena, LiveCodeBench, OpenCompass, SuperCLUE, MMBench, and GAIA.

1.3 Language Model Collection

This index links official sites, APIs, repositories, model hubs, demos, and technical notes for domestic and international language-model ecosystems. It is a starting point for assessing openness, local deployment, context length, reasoning, coding, agent abilities, multilingual support, and API availability.

1.4 Multimodal Model Collection

This directory follows multimodal models for image and video generation, vision understanding, OCR, speech, 3D, digital humans, world models, and device-side deployment. It is useful for comparing capability boundaries and choosing models for creative, document, and interactive applications.

Technical Frameworks

2.0 Embedding Models

Resources for semantic retrieval, vectorization, reranking, and embedding-model selection in RAG and search systems.

2.1 LLM Training Frameworks

Frameworks and practical references for pretraining, fine-tuning, distributed training, data processing, and training-system engineering.

2.2 LLM Inference Frameworks and Deployment

Tools and practices for efficient serving, quantization, acceleration, local deployment, and production inference.

2.3 LLM Evaluation Frameworks

Evaluation systems for quality, safety, task performance, and application-level model validation.

2.4 RLHF

References for preference alignment, reward modeling, reinforcement learning, and post-training methods.

Application Practice

3.0 MCP+A2A

Protocols and practices for connecting AI systems to tools, data, and other agents through MCP and agent-to-agent collaboration.

3.1 RAG + Workflow

Architectures and tooling for retrieval-augmented generation, knowledge workflows, and production RAG applications.

3.2 Agents

Agent frameworks, coding agents, multi-agent collaboration, planning, memory, tool use, evaluation, and real-world automation practices.

3.3 DeepSearch

Resources for deep research, multi-step search, evidence synthesis, and research-oriented agent workflows.

3.4 GraphRAG

Knowledge-graph-enhanced retrieval and reasoning patterns for complex enterprise and research questions.

3.5 NLP2SQL

Methods and products that turn natural-language questions into reliable database queries and data-analysis workflows.

3.6 AI Popular Frameworks

A curated set of commonly used AI development frameworks and engineering references.

Products and Tools

4. AI Applications

Practical AI applications across coding, content creation, productivity, design, research, and vertical domains.

5. AI Products

Product-oriented resources spanning AI MaaS, search, agent products, design, knowledge management, marketing, and digital-human systems.

Learning Resources

6. AI-LLM Competitions

Competition resources for model research, engineering practice, datasets, and community challenges.

6. AI Courses

Structured courses and learning paths for AI fundamentals, large models, and hands-on development.

7. AI Project Recommendations

Curated open-source projects for study, technical investigation, and implementation inspiration.

8.0 Python · 8.1 AI Databases · 8.2 AI Visualization

Core implementation resources for Python, vector and AI databases, visualization, and data-oriented engineering work.

8.3 Machine Learning · 8.4 Computer Vision · 8.5 Recommender Systems

Technical foundations and project references for machine learning, visual intelligence, and recommender systems.

8.6 Reinforcement Learning · 8.7 Knowledge Graphs

Advanced learning resources for decision-making systems, graph computation, knowledge engineering, and graph-based AI applications.

Community and Platforms

9. Learning Platforms · Article Collections · Community Forums

Platforms and communities for continuous learning, technical reading, discussions, open-source participation, and AI ecosystem discovery.

Academic Tools · Interviews · Software

Resources for research, academic productivity, interview preparation, and useful development software.

Enterprise Open Source

10. Datawhale

Community-driven tutorials, open-source projects, and practical learning initiatives.

10. Huawei Open Source · Tencent · Alibaba Open Source

Enterprise open-source ecosystems, including model platforms, agent frameworks, cloud tools, and engineering projects.

10. PaddlePaddle

The PaddlePaddle ecosystem spans deep-learning frameworks, NLP, computer vision, speech, recommendation, model services, and industrial AI practice.

• PaddlePaddle supplementary resources

Future Vision

AI-Compass will continue to follow AI research, products, and engineering practice, while improving the quality and reuse value of its long-term topic pages and weekly highlights.

Follow AI Compass on WeChat

Curated AI updates, practical technical analysis, and implementation cases are published through the project's WeChat channel.

Join AI Compass Knowledge Planet

The Knowledge Planet community provides deeper tutorials, project practice, higher-frequency updates, expert Q&A, and technical discussion.

• AI Compass Knowledge Planet
• AI Compass Knowledge Planet coupon

Technical Blogs CSDN Juejin Zhihu WeChat Official Account Knowledge Planet

Social Media Toutiao Douyin Xiaohongshu

Star History