Mistral Saba 24B Parameter LLM

Mistral is expanding into the Middle East with Saba, a 24-billion-parameter LLM...


Hey there 👋

We hope you're excited to discover what's new and trending in AI, ML, and data science.

Here is your 5-minute pulse...

print("News & Trends")

Image source: Mistral

Mistral is expanding into the Middle East with Saba, a 24-billion-parameter LLM tailored for Arabic. It surpasses its general-purpose model and even supports South Indian languages. With potential regional investors in play, Mistral aims to establish itself as a global AI leader beyond U.S. and Chinese dominance.

Image source: Getty Images

Former OpenAI CTO Mira Murati has founded Thinking Machines Lab, a new AI research and product company built by the minds behind ChatGPT, PyTorch, and Mistral. Their mission? Making AI adaptable, capable, and widely accessible through cutting-edge models, open science, and real-world applications. With a focus on custom AI solutions, foundational advancements, and human-AI collaboration, the lab aims to push AI forward while keeping it practical and transparent.

Image source: Google Cloud

BigQuery ML now supports any open-source LLM from Vertex AI Model Garden — including any models you deploy from Hugging Face and including OSS models you might have tuned. This greatly expands the model choice available to developers. Easily integrate models like Meta Llama 3.3 70B using SQL for tasks like entity extraction and sentiment analysis. The blog provides a tutorial showing how to deploy, query, and analyze real-time insights from unstructured data.

Image source: Alexandre Lallemand/Unsplash

Europe is doubling down on AI sovereignty with OpenEuroLLM, a €37.4M project to build open-source language models for all EU languages. Led by top researchers and backed by EuroHPC supercomputers, the initiative faces skepticism over its sprawling structure and funding. Can it rival corporate AI giants? First models expected by 2026.

Image source: Microsoft

Microsoft just released OmniParser V2, a tool that turns LLMs into powerful computer-use agents, enabling them to analyze and interact with UI elements from screenshots. Paired with GPT-4o, it boosts accuracy from 0.8% to 39.6%, reducing latency by 60%.

print("Applications & Insights")

CASE STUDIES

How Formula 1® uses generative AI to accelerate race-day issue resolution (13 min. read)
Formula 1® partnered with AWS to develop an AI-driven root cause analysis assistant, slashing issue resolution time by 86%. Using Amazon Bedrock, the system automates troubleshooting, reducing manual effort and enabling engineers to resolve critical race-day incidents in minutes instead of days. The AI-powered solution enhances efficiency, reliability, and race experience.

Protecting user data through source code analysis at scale (5 min. read)
Meta’s Anti-Scraping team is using static analysis tools to detect and prevent unauthorized data scraping at scale. By identifying potential vulnerabilities in code before launch, they proactively block scrapers while maintaining user experience. Though not a complete solution, this approach strengthens Meta’s broader strategy to combat data misuse effectively.

The Future of Data: How Decision Intelligence is Revolutionizing Data (4 min. read)
Decision Intelligence (DI) is revolutionizing decision-making by blending AI, data science, and behavioral sciences. Unlike AI, which mimics human intelligence, DI applies it to drive smarter business choices. From retail pricing to healthcare treatments, DI enhances efficiency, reduces risks, and boosts revenue—making it an essential tool for the future of data-driven decision-making.

TUTORIALS

Understanding Probability Distributions for Machine Learning with Python (5 min. read)
Master probability distributions to boost your machine learning models! This guide breaks down key concepts—discrete vs. continuous distributions, common types like Gaussian and Poisson, and how to visualize them in Python. Learn to harness probability for better predictions, data insights, and decision-making in AI.

Retrieval Augmented Generation in SQLite (13 min. read)
Performing retrieval-augmented generation (RAG) doesn’t require complex frameworks or cloud vector databases. This guide walks you through using SQLite with the sqlite-vec extension and OpenAI’s API for embeddings. Learn how to store, query, and retrieve semantically relevant data efficiently—all within a lightweight, local database setup.

print("Research & Advancements")

TOP RESEARCH PAPERS

A Scalable Benchmark for LLMs without Human Judgement (27 min. read)
A new judge-free benchmark evaluates LLM open-ended generation using n-gram statistics and rule-based metrics—Fluency, Truthfulness, and Helpfulness—without relying on human or AI judges. The approach correlates well with GPT-4o assessments while being computationally efficient, offering a scalable alternative for evaluating LLMs.

Speeding up LLM Inference with CopySpec (30 min. read)
CopySpec accelerates LLMs by detecting repeated text patterns and copying them instead of regenerating each token. It quickly verifies copied text and fills in only missing parts, reducing computation without extra GPU memory. This boosts speed up to 3.08× and enhances speculative decoding by 49%, keeping responses fast and accurate.

TOP REPOSITORIES

AI Search
zaidmukaddam/scira
☆ 6.3K stars
Scira (formerly MiniPerplx) is minimalistic open-source AI search engine designed to streamline your internet queries. Powered by the Vercel AI SDK, Scira integrates advanced models like Grok 2.0 and Claude 3.5 Sonnet, to deliver precise and efficient search results. Experience a smarter way to find information online.

MLOps
langgenius/dify
☆ 69.5K stars
Dify is an open-source platform that simplifies building AI applications by integrating Backend-as-a-Service and LLMOps. It offers features like visual orchestration, RAG pipelines, agent capabilities, and model management, enabling rapid prototyping to production. Dify supports various LLMs, including OpenAI's GPT series and open-source models like Llama2. Users can deploy applications via Docker or Kubernetes and access a cloud service with free GPT-3.5 requests.

Text Processing
microsoft/markitdown
☆ 38.1K stars
Microsoft's MarkItDown is a Python tool that transforms various file types—including PDFs, Word documents, and images—into Markdown format. Ideal for indexing and text analysis, it supports a wide range of formats and offers a plugin-based architecture for extensibility. Easily integrate it into your workflow to streamline content conversion.

Computer Vision
microsoft/OmniParser
☆ 13.6K stars
Microsoft's OmniParser is an open-source tool that converts UI screenshots into structured, easy-to-understand elements. This enhances AI models' ability to generate context-aware actions. OmniParser supports various LLMs and offers easy deployment via Docker or Kubernetes. Experience smarter screen parsing today.

That’s it for today!

Before you go we’d love to know what you thought of today's newsletter to help us improve the pulse experience for you.

What did you think of today's pulse?

Your feedback helps me create better emails for you!

Login or Subscribe to participate in polls.

See you soon,

Andres