AI Trending Models Skill#

Purpose#

Systematically collect and summarize the latest, most viral open-source AI / large-language-model (LLM) projects from authoritative sources, and present a structured, actionable intelligence report to the user.

Trigger Conditions#

Load this skill whenever the user asks to:

• Find / collect / scout the latest trending AI open-source projects
• Summarize recent hot LLM / multimodal / agent frameworks
• Track what's blowing up on GitHub, HuggingFace, arXiv, or AI news outlets

Workflow#

Step 1 — Run the automated fetcher (preferred)#

Execute scripts/fetch_trending.py to pull live data from multiple sources:

bash

python3 scripts/fetch_trending.py

The script outputs a structured JSON file ( trending_report.json) with raw data. Read and interpret that file for the final report.

If the script cannot run (network issues, missing deps), fall back to Step 2.

Step 2 — Manual web research fallback#

Query each source listed in references/sources.md using web_fetch or web_search. Collect at minimum:

• GitHub Trending (past 7 days, filter: AI / ML / LLM)
• HuggingFace Models — trending tab
• arXiv cs.AI / cs.CL — last 7 days, sorted by submission count
• Papers With Code — trending methods
• Twitter / X — #OpenSourceAI, #LLM hashtags top posts

Step 3 — Deduplicate & rank#

Rank projects by composite signal:

1. GitHub stars velocity (stars gained / days since release)
2. Cross-source mention frequency (appears in GitHub + HuggingFace + arXiv = higher rank)
3. Recency (prefer projects released or updated within 30 days)
4. Community buzz (forks, issues, PR activity, social mentions)

Step 4 — Produce the report#

Output a clean Markdown report following the template in references/report_template.md.

Key sections:

• 执行摘要 / Executive Summary — 3-sentence overview of what's hot right now
• TOP 10 爆火项目 — ranked table with: rank, project name, org/author, stars ⭐, stars delta Δ, category, one-line description, link
• 按方向分类 — group projects by: LLM底座 | 多模态 | Agent/工具链 | 推理加速 | 数据/微调 | 其他
• 值得关注的论文 — top 5 arXiv papers linked to open-source code
• 趋势洞察 — bullet-point analysis of what the data signals for the industry

Output Standards#

• Language: match the user's language (default Chinese 中文)
• Format: Markdown with emoji for visual clarity
• Stars count: use K notation (e.g. 12.4K)
• Always include direct URLs to GitHub repos / HuggingFace model pages
• Date-stamp the report header with the collection date
• If data is older than 3 days, note it clearly

Quality Rules#

• Minimum 10 projects in the main table; aim for 15–20
• No duplicates across GitHub / HuggingFace entries for the same project
• Verify each project is genuinely open-source (has an open license)
• Flag projects that are "demo-only" or have no released weights
• Prioritize projects with working code over paper-only releases