12th March 2025

Create and deploy LLM agents just using natural language!🔥

AutoAgent is the Fully-Automated & Zero-CodeLLM Agent Framework that let’s you create and deploy LLM agents using just natural language.

Key Features:

📚 Agentic-RAG – Built-in self-managing vector database, outperforming LangChain.

✨ Zero-Code Agent & Workflow Creation – Just use natural language, no coding needed.

🌐 Universal LLM Support – Works with OpenAI, Anthropic, Deepseek, vLLM, Huggingface & more.

🔀 Flexible Interaction – Supports both function-calling & ReAct modes.

The best part?

It’s 100% Open Source

Github Repo: https://github.com/HKUDS/AutoAgent

Announcement from Takween AI

We’re excited to announce a strategic partnership between Takween AI and Groq to accelerate AI adoption in Saudi Arabia! 🇸🇦✨

By combining Groq’s ultra-fast AI inferencing infrastructure with Takween AI’s cutting-edge solutions, we’re bringing seamless, high-performance AI capabilities to businesses across the Kingdom. This partnership will:

🔹 Expand AI accessibility and innovation across industries

🔹 Empower businesses with cutting-edge AI-driven solutions

🔹 Support Saudi Vision 2030’s digital transformation goals

🤝 This agreement was signed by Fahad AlTurief VP & MENA MD of Groq and Ahmed Sulaiman Sulaiman, Founder & CEO of Takween, marking a significant milestone in AI innovation.

Together, we’re shaping the future of AI in Saudi Arabia, driving innovation, and unlocking new opportunities for growth! 🚀📍

🔥AI Auctions? How Token Bidding Could Reshape LLM Collaboration

New research introduces token auctions, a mechanism where multiple LLMs collaborate by bidding on each word to generate a joint output.

This method ensures fairer, more optimized AI-generated content in scenarios where multiple stakeholders have competing interests.

✅How it works – Each LLM submits bids for the next token, and an auction mechanism selects the final output based on preferences and strategic bidding.

✅Why it matters – This approach balances competing AI inputs, making it ideal for applications like co-authored reports, ad generation, or content aggregation.

✅Proven results – Tested with LLM-generated ad creatives, the model successfully merged competing advertiser inputs into a unified message.

MANUS AI: HYPE VS. REALITY 🔍

Manus AI has been making waves on social media with impressive demos, but it is not “China’s next DeepSeek moment”. The co-founder confirmed details on X:

✅ Built on Anthropic Claude Sonnet, not their own foundation model

✅ Has access to 29 tools and uses @browser_use open-source for browser control

✅ User communicates with executor agent and not planner or other agents.

✅ Each user gets isolated sandbox environment

✅ Outperforms OpenAI Deep Research on GAIA benchmark

Tools and Prompts: https://gist.github.com/jlia0/db0a9695b3ca7609c9b1a08dcbf872c9

Building AI products doesn’t require training your own foundation models. We’re probably just scratching the surface of what existing models can do with the right tooling and integration!

🚀 Claude 3.5 & GPT-4o Get a Web Crawling Boost

A new tool, Firecrawl, is making it easier to convert entire websites into LLM-ready markdown or structured data.

1️⃣Works with Claude 3.5 and GPT-4o to scrape, crawl, and extract data with just a few lines of code.

2️⃣Captures all accessible subpages—no sitemap needed.

3️⃣ Outputs clean, structured data optimized for LLM-based workflows.

This approach streamlines data collection for AI applications, making web content more accessible for large-scale models.

How do you see this impacting AI-driven data extraction? 🤔

Source: Eric Vyacheslav

EuroBERT is released

A multilingual encoder model family (210M to 2.1B parameters) trained on 5T tokens across 15 languages, with support for sequences up to 8,192 tokens. It’s open-source and designed to power multilingual retrieval, classification, and embeddings.

🔹 Why EuroBERT?

✅ State-of-the-art performance across multilingual retrieval, classification, and regression

✅ Long-context support (8,192 tokens) for document-level understanding

✅ Mathematics & Code training for improved reasoning

✅ Outperforms XLM-RoBERTa, mGTE, and other leading models

EuroBERT builds upon our team’s experience training EuroLLM & CroissantLLM, but encoders are NOT decoders and require specific design decisions! We ran extensive ablations on masking ratios, language and data distributions, annealing, and data quality to ensure optimal performance.

📢 Beyond the results we report, nothing beats people fine-tuning the model for their usecases and sharing real-world feedback, so feel free to do so, everything is on the HuggingFace page !

More details and links in the blog: https://huggingface.co/blog/EuroBERT/release

New Agent Leaderboard!

Hugging Face created a Leaderboard for smolagents. It evaluates open and closed model on a mini-subset of GAIA, MATH, SimpleQA. 👀

🧠 GAIA evaluates General AI Assistants on real-world tasks that are easy for humans but challenging for AI.

🧮 MATH evaluates the mathematical reasoning capabilities from high school math competitions, covering topics like algebra, geometry, probability, and calculus.

🔍 SimpleQA evaluates the ability to accurately answer short, fact-seeking questions.

Leaderboard: https://huggingface.co/spaces/smolagents/smolagents-leaderboard

Code: https://github.com/huggingface/smolagents/tree/main/examples/smolagents_benchmark