Skip to main content
Back

Mistral AI Drops Open-Souce 24B Models, Beats Gemma 3 and GPT-4o Mini

TRY NOW

March 18, 2025 6 min read

Hey,

Welcome to AlphaSignal – the most read newsletter by AI developers. 


We bring you the top 1% of news, papers, models, and repos, all summarized to keep you updated on the latest in AI.

IN TODAY'S SIGNAL

Read time: 6 min 30 sec

🎖️ Top News

📌  Augment

⚡️ Trending Signals

📌  AI21


💻  Top Repo

  • awesome-mcp-servers: a curated list of MCP servers for AI model-resource interaction with frameworks, tutorials, and tips.


  • camel: a framework for studying multi-agent scaling laws, supporting various agents, tasks, models, and environments.

  • public-apis: a curated list of public APIs across various domains for easy integration.


🧠  Top Lecture

If you're enjoying AlphaSignal please forward this email to a colleague. 

It helps us keep this content free.

TOP NEWS

AI Model

Mistral AI unveils Mistral Small 3.1, a lightweight 24B open-source model outpacing Gemma 3 and GPT-4o

⇧ 9,889 Likes

What's New

French AI lab, Mistral AI has released Mistral Small 3.1, a 24-billion-parameter open-source model with an expanded 128k token context window. The model features an expanded context window of up to 128k tokens and delivers fast inference speeds. Mistral Small 3.1 outperforms competitors like Gemma 3 (27B) and GPT-4o Mini across several benchmarks.


Key Highlights

  • Expanded context window of 128k tokens for handling larger inputs in a single pass.

  • Inference speed of 150 tokens per second for low-latency performance.

  • Runs on accessible hardware, including RTX 4090 and Macs with 32GB RAM.

  • Supports fine-tuning for specialized domains such as legal, medical, and technical applications.

Performance Benchmarks

Mistral Small 3.1 outperforms competing models in text generation and multimodal tasks.

  • GPQA Main: 44.42%, surpassing Gemma 3-it (36.83%) and GPT-4o Mini (40.2%). Measures general knowledge question accuracy.

  • MMLU: 80.62%, ahead of Gemma 3-it (76.9%) and Claude-3.5 Haiku (77.6%). Assesses general understanding and reasoning ability across various topics.

  • HumanEval: Matches GPT-4o Mini and Claude-3.5 Haiku. ests the model’s ability to solve coding problems through Python scripts.

  • TriviaQA: Matches Gemma 3-it (27B). Evaluates the model's performance in answering trivia questions.


Multimodal Capabilities
The model processes both text and images for advanced understanding.

  • Supports document verification, diagnostics, and quality control.

  • Enhances image-based customer support and real-time decision-making.

  • Enables object detection in security applications.


Optimized for On-Device Use
Mistral Small 3.1 runs on accessible hardware, such as RTX 4090 and Macs with 32GB RAM.

  • Optimized for both consumer and enterprise deployments.

  • Supports fast-response tasks like virtual assistants and automated workflows.

  • Lightweight enough for on-device use, making it efficient for small-scale applications.

Fine-Tuning and Customization
You can fine-tune the model for domain-specific tasks.

  • Supports LoRA and full fine-tuning.

  • Provides base and instruct checkpoints for further customization.

  • Optimized for legal, medical, and technical applications.

Deployment and Availability
Mistral Small 3.1 is available for download and cloud-based access.

  • Released under an Apache 2.0 license.

  • Available on Hugging Face, Mistral AI’s La Plateforme, and Google Cloud Vertex AI.

  • Coming soon to NVIDIA NIM and Microsoft Azure AI Foundry.

Community Feedback

Expert Rob
"24B parameters outperforming larger models while maintaining sub-12ms latency? That's some serious architecture optimization."


Thomas Jebarsan
"
The advancements in Mistral AI Small 3.1 are truly impressive, especially with its focus on multimodal and multilingual capabilities. The low-latency function calling feature is particularly intriguing for real-time applications."


cedric
"Great local model. They mentioned a 32GB MacBook, but still no quantized GGUF or MLX releases, surprising since they partnered with Ollama for Mistral Small 3. Expect quantized models to start popping up from the community soon!"

Try now

Need an AI Assistant Built for Real-World Code?

Augment Code is designed for pro engineers tackling complex, production-grade codebases.

  • Handles massive, intricate repositories with ease.

  • Works seamlessly with Vim, VSCode, JetBrains, and more.
  • Understands your dependencies and team’s coding standards.

Skip the gimmicks. Build with Augment Code.

Try now

partner with us →

TRENDING SIGNALS

Speech Synthesis

Sesame releases CSM-1B, a 1B parameter model of its near-human speech model with Apache 2.0 license for commercial use

⇧ 7,495 Likes

AI Tool

AI2 announces a 32B parameter open model outperforming GPT-3.5 and GPT-4o mini, offering fully open access to data and code

⇧ 2,017 Likes

Generative AI

Roblox introduces an open-source AI model for generating 3D objects and scenes from text; API in Roblox Studio also available

⇧ 2,748 Likes

AI Assistant 

Google updates Gemini with personalized AI responses, integrating Search history to understand your activities and preferences

⇧ 1,147 Likes

AI Security

Fiddler AI and AWS presents a webinar discussing AI agent governance, security risks and best practices for agentic workflows

⇧ 649 Like

Deploy Anywhere with Confidence: Flexible LLM for Enterprise

Jamba 1.6 is an open-weight AI model built for enterprise integration.

  • 256K Context Window – Handles long documents and conversations efficiently.

  • Flexible Deployment – AI21 SaaS or self-host for full data control.

  • Optimized for Speed – Jamba Mini 1.6 ensures low-latency responses.

  • Hybrid Architecture – Mamba-Transformer MoE design improves efficiency and cost-effectiveness.

Deploy Jamba 1.6 today ↗️

TRENDING REPOS

MCP

awesome-mcp-servers

☆ 11,393 Stars

This repo helps you explore curated Model Context Protocol (MCP) servers for efficient AI model-resource interaction. It includes production-ready and experimental servers, supporting file access, database connections, and API integrations. The list offers useful frameworks, tutorials, and tips to optimize model deployment and resource utilization in real-world environments.

Agents

camel

☆ 10,614 Stars

It helps you study the scaling laws of AI agents by providing an open-source multi-agent framework. It includes agent implementations, tasks, prompts, models, and simulated environments. It supports large-scale experiments to analyze agent behaviors, capabilities, and risks. The framework facilitates research on agent interactions and emergent properties in complex settings.

API

public-apis

☆ 333,082 Stars

This repo helps you find and use public APIs across various domains. It provides a manually curated list of free APIs for integration into applications. Categories include finance, machine learning, weather, and more. Community members maintain the list to ensure reliability. It serves as a central resource for discovering useful APIs.

TOP LECTURE

Normalization in Deep Learning

Optimizing Transformers: How DeepSeek R1 Improves Speed and Efficiency

⇧ 17,483 Likes

The "How DeepSeek Rewrote the Transformer" lecture explains the architecture and optimizations behind DeepSeek R1, a language model that improves Transformer efficiency.

It details Multi-Head Latent Attention (MHLA), a novel method that reduces memory usage while increasing inference speed. DeepSeek R1 achieves a 57x reduction in KV cache size and 6x faster text generation compared to traditional Transformers.

You will learn:

  • How Transformers use attention to process token relationships.

  • The mechanics of Key-Value (KV) caching and its impact on performance.

  • How Multi-Head Latent Attention (MHLA) optimizes KV caching.

  • The difference between multi-query attention, grouped query attention, and MHLA.

  • How DeepSeek's implementation improves efficiency while maintaining model accuracy.
Watch now

At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.


Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.

Looking to promote your company, product, service, or event to 200,000+ AI developers? Let's work together.

Work with us