Welcome back to Daily AI Skills.

Here’s what we are covering today:
1. Upgraded DeepSeek V3 - The Best Non-Thinking LLM
2. Reve - The Stealth Startup that’s beating Midjourney
3. Gemini 2.5 - Google’s Most Intelligent AI Model

+ The LLM Engineer Toolkit

The Best Non-Reasoning Model?

Chinese AI startup DeepSeek has unveiled an updated version of its V3 model, a 641GB model designed to run on high-end personal computers. This latest release also features a highly permissive MIT open-source license, allowing for broad usage.

The details:

• The V3-0324 update incorporates a Mixture-of-Experts architecture, activating only 37B parameters per token, significantly reducing computational demands.

• Testers have confirmed that it runs smoothly on Apple’s Mac Studio, making it the first model of its size and capability to be accessible outside traditional data centres.

• Early users report enhanced math and coding performance, with one describing it as the best non-reasoning AI model available.

• Unlike the previous version, which had a more restrictive custom license, V3-0324 is now available under an MIT open-source license for greater flexibility.

Read the Model Card on Hugging Face here

Try out the model on web here: https://platform.deepseek.com/

Did a Stealth Startup just beat MidJourney?

Reve has officially emerged from stealth with Reve Image 1.0, a new text-to-image AI model that recently dominated global rankings under the codename “Halfmoon”—demonstrating outstanding prompt accuracy, text rendering, and image quality.

The details:

• The model secured the #1 spot in Artificial Analysis' Image Arena, surpassing competitors like Google’s Imagen 3, Midjourney v6.1, and Recraft V3.

• Reve’s goal is to integrate logic into visual generative models, with 1.0 excelling in prompt fidelity and long text rendering during evaluations.

• The platform includes natural language editing, photo uploads, and an ‘explore’ tab where users can browse community-generated images.

• Reve Image 1.0 is currently available for free preview, though API access is not yet available. The company hints that more features are on the way.

Try out the model here: https://preview.reve.art/app

Gemini 2.5 - Google’s Most Intelligent AI Model

Google DeepMind has introduced Gemini 2.5, its most advanced AI model yet, featuring groundbreaking improvements in reasoning, coding, and multimodal capabilities

The details:

• The first release, Gemini 2.5 Pro Experimental, outperforms competitors, debuting at #1 on LMArena and excelling in coding, math, and science benchmarks.

• Enhanced reasoning capabilities allow the model to analyze information more effectively, scoring state-of-the-art results in rigorous human knowledge tests like Humanity’s Last Exam (18.8%).

• Advanced coding performance includes major improvements over Gemini 2.0, enabling web app creation, agentic coding, and code transformation. On SWE-Bench Verified, it achieves a 63.8% score with a custom agent setup.

• Native multimodality and an expanded 1 million token context window (with 2 million coming soon) make 2.5 Pro highly adept at handling text, audio, images, video, and code repositories.

• Availability: Developers can experiment with Gemini 2.5 Pro in Google AI Studio, while Gemini Advanced users can access it via desktop and mobile. It’s set to arrive on Vertex AI in the coming weeks.

Read the full article by Google here

The LLM Engineer Toolkit

📩Forward it to people you know who are keeping pace with the changing AI world, and stay tuned for the next edition to stay ahead of the curve!

Maybe you missed out on these:

1. Copilot vs Claude vs Gemini

2. How I Built an AI Influencer in 2 Minutes—and Landed Sponsorships

3. Microsoft Muse, Helix and Google's AI Co-Scientist

4. Bolt vs Lovable

5. Your AI Stack Just Got Superpowers: 4 New Inference Providers Live!