- The Rundown AI
- Posts
- Stealth startup dethrones image giants
Stealth startup dethrones image giants
PLUS: DeepSeek quietly drops V3 model upgrade
Rowan Cheung
March 25, 2025
Good morning, AI enthusiasts. The text-to-image race just got a surprising new frontrunner — with Reve emerging from stealth with a model that instantly topped global rankings under the codename "Halfmoon."
With next-level design and text capabilities, indistinguishable-from-reality photorealism, and world-class prompt adherence, a new AI creative powerhouse just announced itself in a BIG way.
In today’s AI rundown:
Reve’s new leading image model
DeepSeek quietly drops V3 upgrade
Turn any YouTube video into your tutor
ARC Prize returns to challenge AI reasoning
4 new AI tools & 4 job opportunities
LATEST DEVELOPMENTS
REVE
Image source: Reve / The Rundown
The Rundown: Reve just emerged from stealth with Reve Image 1.0, a new text-to-image AI model that topped global rankings with the codename “Halfmoon” over the last week—showcasing exceptional prompt accuracy, text rendering, and image quality.
The details:
The model claimed the #1 position in Artificial Analysis' Image Arena, outperforming rivals like Google's Imagen 3, Midjourney v6.1, and Recraft V3.
Reve said its mission is to “enhance visual generative models with logic,” with 1.0 showing impressive prompt adherence and long text rendering in tests.
The platform also features natural language editing, photo uploads, and an ‘explore’ tab to view community prompts and generations.
A preview of Reve Image 1.0 is currently free to try (though no API access yet), with the company saying that “much more is coming soon”.
Why it matters: What a stealth debut from Reve, with their first model already topping the leaderboards against established giants in the text-to-image arena. 1.0 seems to combine the best of the SOTA image models — with extreme photorealism, world-class prompt following, editing tools, and absolutely next-level text capabilities.
TOGETHER WITH VANTA
The Rundown: Navigating new compliance requirements can be a daunting task, but with the right automation tools, it doesn't have to be. Whether you’re a fast-growing startup or an established security team, Vanta can help you achieve continuous compliance (and more).
Join the live demo on April 3 to learn how Vanta can help you:
Automate compliance for frameworks like SOC 2, ISO 27001, HIPAA, HITRUST, and ISO 42001
Continuously monitor controls to build and reinforce your security foundation
Scale your compliance program and manage vendor risk with automated tools and AI
DEEPSEEK
Image source: Deepseek
The Rundown: Chinese AI startup DeepSeek just released an updated version of its V3 model, a massive 641GB model capable of running on high-end personal computers — also featuring a highly permissive open source MIT License for broad use.
The details:
The V3 update, V3-0324, uses a Mixture-of-Experts architecture that activates only 37B parameters per token, dramatically reducing compute demands.
Testers have shown it can run smoothly on Apple's Mac Studio computers, making it the first model of this caliber accessible outside data centers.
Early users have also reported upgraded math and coding capabilities, with another calling it the best non-reasoning model available.
V3-0324 can be accessed with an open-source MIT License, a change from the previous V3 model that had a more restrictive custom license.
Why it matters: China’s AI darling continues to ship, with a supposedly minor update bringing some big upgrades. Rumors about the upcoming R2 release are also gaining momentum, hinting at another 'DeepSeek moment' that could shake the AI world—potentially signaling a new leader in the field.
AI TRAINING
The Rundown: In this tutorial, you will learn how to convert YouTube videos into personal tutors with Google AI Studio, which lets you ask questions about any video by simply pasting the link—making complex content instantly accessible for learning.
Step-by-step:
Visit Google AI Studio and log in with your Google account.
Select "Gemini 2.0 Flash" from the Model dropdown menu on the right side of the screen.
Paste your YouTube video link in the prompt area, followed by your specific question about the content.
Ask follow-up questions to explore the video content more deeply, referencing specific timestamps if needed.
Pro tip: When exploring videos on complex educational subjects, you can ask the AI to "Create a simple explanation of the [concept]" or "Generate a quiz based on this lecture to test my understanding."
PRESENTED BY INNOVATING WITH AI
The Rundown: Innovating with AI's new program, AI Consultancy Project, transforms AI enthusiasts into professional consultants — tapping into a market projected to reach $54.7B by 2032.
The 6-month program delivers:
Proven frameworks for client acquisition and service delivery
A step-by-step path to six-figure consulting income
Students who land their first AI client in as little as 3 days
ARC FOUNDATION
Image source: ARC Foundation
The Rundown: The ARC Prize Foundation just launched ARC-AGI-2, a new benchmark designed to push the frontier of AI reasoning capabilities — alongside a $1M competition to drive research toward more efficient general intelligence systems.
The details:
ARC-AGI-2 targets skills that remain difficult for AI while being easy for humans, with tasks solvable by at least two humans in under two attempts.
Current AI reasoners perform poorly on ARC-AGI-2, with even OpenAI's o3-low scoring only an estimated 4%, compared to 75.7% on the previous version.
ARC also introduced an efficiency metric alongside raw performance, measuring cost per task to test both capability and resource efficiency.
The ARC Prize 2025 competition launches this week with $1M in prizes, including $700k for the first to achieve 85% accuracy within efficiency limits.
Why it matters: Many of the top AI labs feel that AGI is just around the corner, and ARC’s original AGI-1 benchmark was already seeing top models like the o3 score in the 80s (not without controversy). While the new iteration may be a better test, it’s still hard to imagine it lasting long — but the AGI goalposts will likely continue to shift.
QUICK HITS
🪄 Promptimize - Browser extension that gives you the power of an expert prompt engineer in a single click*
🧠 Hunyuan T1 - Tencent AI matching DeepSeek in performance and pricing
⚙️ Zapier MCP - Connect AI across 8,000+ apps without API integrations
🤔 Claude “Think” - Tool to improve Claude’s problem-solving performance
*Sponsored Listing
Sam Altman announced new internal leadership promotions at OpenAI, with Mark Chen appointed as Chief Research Officer and Brad Lightcap’s role expanding as COO.
Alibaba’s Qwen team open-sourced Qwen2.5-VL-32B-Instruct, a new vision-language model featuring enhanced mathematical reasoning and visual capabilities.
Netflix co-founder Reed Hastings donated $50M to his alma mater, establishing the Hastings Initiative for AI and Humanity to study the tech’s risks and future impact.
Alibaba-affiliate Ant Group reportedly adopted a hybrid chip strategy combining Chinese and American chips, which reduced AI development costs by 20%.
Google started rolling out ‘Project Astra’ features that give advanced vision, live video, and screen reading capabilities to Gemini.
Alibaba released LHM, a new AI model that creates animated 3D avatars from just a single reference image.
COMMUNITY
Join our next workshop this Friday at 4 PM EST with Dr. Alvaro Cintas, The Rundown’s AI professor. You'll learn how to use the latest AI tools to take your content creation to the next level.
RSVP here. Not a member? Join The Rundown University on a 14-day free trial.
We’ll always keep this newsletter 100% free. To support our work, consider sharing The Rundown with your friends, and we’ll send you more free goodies.
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 Rundown experience for you. |
See you soon,
Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team
Reply