What moved in AI this week: models, apps and the tools people build with. It's built automatically every Monday from public numbers: what people read about on Wikipedia, what developers install, and what gets attention on Hacker News, GitHub and Hugging Face.
DeepSeek added a web search API for agents, a compact open multimodal model trended, and installs of the MCP software kits kept growing.
DeepSeek adds search and page-fetch APIs for AI agents
DeepSeek published two web APIs: one returns search results from its own search engine, the other turns a web page into Markdown text. They are meant for agents and apps that need live web content. The docs say failed, empty or timed-out requests are not charged. The announcement drew 566 points on Hacker News.
MiniCPM-V 4.7, a compact open multimodal model, is trending
OpenBMB's MiniCPM-V 4.7 is an open model that reads images and video. It has 35 billion parameters in total; the 'A3B' in its name means about 3 billion are active for each token (a mixture-of-experts design), which keeps it light enough to run on phones. It is released under the Apache-2.0 license and ranks #2 on Hugging Face trending.
MCP (Model Context Protocol) is a standard way for AI apps to connect to outside tools and data. Installs of the TypeScript SDK grew 28% and the Python SDK 14%, level with FastMCP at 14%. The TypeScript SDK shipped 1.32.0 on 2 October, which limits when its HTTP client follows redirects.
Change in weekly English Wikipedia views, for pages with at least 5,000 views a week.
Vibe coding+32%
DeepSeek+28%
Model Context Protocol+25%
Mistral AI+18%
Qwen+12%
Perplexity−7%
Views by people (not bots) of each English Wikipedia article last week: a rough measure of how many people looked something up. The trend compares the median of recent weeks with the weeks before, and a dashed line means a spike, usually a news story.
Install trend for the six biggest risers and four biggest fallers with at least 10,000 installs a week.
faster-whisper+63%
Ragas+41%
Inspect+37%
Mastra+34%
DeepEval+32%
Vercel AI SDK+31%
TRL−11%
Unsloth−15%
Mistral SDK−16%
Weaviate−45%
Installs last week from PyPI and npm. The trend compares the median of recent weeks with the weeks before. A dashed line means some weeks had bot-sized spikes. This week, PyPI trends use only the last 7 weeks, because the registry changed how it counts.
opencode installs (+30%): Fastest-growing coding agent on npm, but no release or event found to explain it.
Suno on Wikipedia (+171%): Views climb from a low base with a one-week spike and no event to check.
Character.ai on Wikipedia: One-week jump in views with no product release to check against a primary source.
SGLang and Weaviate installs: Large falls (-62%, -45%) with no release or event to explain them.
Gemini 4 Argon: Already picked last week; nothing new in the data.
OpenAI engineer arrested: News about people, not about AI tools.
how this page is made
A Claude Code job runs every Monday. A script collects the numbers: Wikipedia page views from Wikimedia, weekly installs from pypistats.org and the npm registry, stars and releases from GitHub, trending models and papers from Hugging Face, and the top AI stories on Hacker News. The agent never writes a number. It picks what mattered this week and writes the notes. A second agent checks each note against its sources and drops anything it can't confirm. A last script checks that every number in the notes matches the data and that every link works, and the page is published only if all of that passes. It has run 2 times so far: 2 published. The last run, on 19 Oct, published.
Page views measure curiosity, not use, and install counts include CI and mirrors, so read both as trends, not user counts. Raw data for this week