Frontier AI Update
Bigger Contexts, Faster Models, Smarter Agents
⚡ Model Progress
1、Grok 4 Fast now has a huge 2M token context window. I use Gemini more often because it handles long contexts well, but now there’s a new option available.
Moreover, reasoning ability has also significantly improved, from 77.5% to 94.1%. This percentage is hard to believe until you try it yourself.
2、Apple has released FastVLM and MobileCLIP2 on Hugging Face.
These models deliver up to 85x faster processing speeds and are 3.4x smaller than previous models, enabling real-time VLM applications!
They even support real-time video captioning entirely within the browser (no installation required). This is a significant step forward for accessibility!
💻 Tools and Platforms
1. ollama now supports running Kimi K2 Thinking via `ollama run kimi-k2-thinking:cloud`. Kimi K2 Thinking is a
open-source thinking agent designed for stepwise reasoning and dynamic tool invocation.
Exceptional multi-step reasoning capabilities: It excels in multi-step reasoning and tool usage, outperforming or matching Sonnet 4.5 and GPT-5 on 𝜏²-Bench and GPQA Diamond benchmarks.
2、Gemini API launches a document search tool that can directly interact with PDFs. Developers can also use it to build their own knowledge base assistants, with Gemini providing free storage and free vector generation during queries.
demo app: https://aistudio.google.com/apps/bundled/ask_the_manual
3、Perflexity’s Comet Assistant has recently been upgraded. This assistant can automatically search for information online and complete tasks for you. For example, it can help me create a spreadsheet to track tasks, search websites to find the lowest ticket prices, and look for job openings.
4、OpenAI sora APP roll out character cameos
5、You can use application chat in ChatGPT.
6、Copilot can handle existing pull requests, so you can tag it in any manually created pull request and ask it to make changes.
🧩 Agent
1、AndrewYNg’s Agent Build Course : https://www.deeplearning.ai/courses/agentic-ai/
2、Claude Agent SDK Loop is an agent framework for building various AI agents.
3、Google Launches: Data Formulator v0.5: Vibe with data, in control Analysts leverage AI agents to directly converse with data and gain insights
💡 Trends and Perspectives
1. Google released its third-quarter earnings report. Revenue surpassed $100 billion for the first time in a single quarter, driven primarily by double-digit growth across all major business segments. (Quarterly revenue stood at $50 billion just five years ago.) The full-stack AI approach is fueling genuine momentum and rapidly delivering products.
Google is making too much money!
If you’re finding this newsletter valuable, share it with a friend, and consider subscribing if you haven’t already.
Sincerely,
Felix 👋