Use AI as Your Personal Assistant’

Ready to save precious time and let AI do the heavy lifting?

Save time and simplify your unique workflow with HubSpot’s highly anticipated AI Playbook—your guide to smarter processes and effortless productivity.

Download the free guide today.

Today:

• DeepSeek Releases Prover V2 Model

• Qwen Releases Compact Multimodal Model

• Google Eyes 2025 iPhone Integration

• OpenAI Fixes Overly Flattering ChatGPT

• Gemini Adds Native Image Editing

DeepSeek Releases Prover V2 Model

Chinese AI lab DeepSeek has quietly released version 2 of Prover, its math-focused AI model, on Hugging Face. Built on the company’s massive V3 model with 671 billion parameters, Prover V2 uses a modular design that assigns tasks to specialized components. The upgrade strengthens its ability to solve mathematical proofs. DeepSeek also plans to update its reasoning model soon, signaling continued advances in AI-driven logic and problem-solving.

Why This Matters

1. Advances Formal Reasoning: Prover boosts AI's ability to solve complex math and logic problems, a key step toward advanced scientific applications.
   

2. Open Access: Releasing it on Hugging Face enables broader research collaboration and transparency.
   

3. China’s Growing AI Momentum: It reflects China’s rising presence in cutting-edge AI, especially in high-skill domains like theorem proving.

Qwen Releases Compact Multimodal Model

Alibaba’s Qwen team launched Qwen2.5-Omni-3B, a powerful yet compact AI model that handles text, audio, image, and video on consumer-grade PCs. Despite being smaller, it delivers over 90% of the performance of its 7B counterpart and supports real-time speech and multimodal input. While technically impressive, it's licensed for research only—companies must get special permission to use it commercially, making it a strong evaluation tool, not a deployment-ready solution.

Why This Matters

1. Multimodal Power on Consumer Hardware: It shows that real-time, high-quality multimodal AI can now run outside data centers.
   

2. Lower Barrier to Experimentation: Developers can test and build with advanced models without expensive hardware.
   

3. Strategic Licensing Signals Shift: The research-only license highlights a growing trend where open models are technically open but commercially gated.

Google Eyes 2025 iPhone Integration

Google CEO Sundar Pichai told a U.S. court that he hopes to make Gemini a built-in option on iPhones by the end of 2025. Apple currently relies on its own models and OpenAI’s ChatGPT for AI features across its devices. If successful, the deal would place Google’s Gemini alongside ChatGPT in Apple’s ecosystem, deepening competition between AI giants for mobile integration dominance.

Why This Matters

1. Major Distribution Boost: Being built into iPhones would massively expand Gemini’s user base.
   

2. AI Platform Wars: The move intensifies competition between Google and OpenAI for consumer reach.
   

3. Shifts in Apple Strategy: Apple’s openness to third-party AI marks a significant shift in its tightly controlled ecosystem.
   

🧠RESEARCH

Reinforcement Learning for Reasoning in Large Language Models with One Training Example

This paper shows that training a language model with just one carefully chosen math example using reinforcement learning can dramatically boost performance. The technique nearly doubles accuracy on key benchmarks and works across models. It also uncovers new behaviors like cross-domain gains and improvements even after training accuracy peaks, challenging assumptions about model learning.

UniversalRAG: Retrieval-Augmented Generation over Multiple Corpora with Diverse Modalities and Granularities

UniversalRAG is a new system that improves how AI models pull in outside information to answer questions. Unlike typical methods that rely on just text or one type of data, it retrieves from multiple sources—text, images, videos—based on what fits best. This smarter, targeted retrieval boosts accuracy across diverse tasks.

ReasonIR: Training Retrievers for Reasoning Tasks

ReasonIR-8B is a new retriever built specifically for reasoning tasks, not just simple fact-finding. It’s trained on synthetic queries paired with tricky false leads, making it better at picking helpful information. It outperforms other systems on complex benchmarks and boosts performance in tasks like MMLU and GPQA. Code and models are open source.

🛠️TOP TOOLS

PicFinder AI - AI-powered image generation tool that transforms textual descriptions into unique visual content.

Unboring AI - AI-powered online platform that offers innovative tools for face swapping, photo animation, and video restyling, allowing users to create engaging and entertaining content with just a few clicks.

Crayo - Transform simple text inputs into viral-ready videos.

Dubverse AI - AI-powered platform that revolutionizes video dubbing and subtitling, offering creators and businesses the ability to make their content multilingual instantly across 30+ languages with human-like AI voices.

Bing Create - AI-powered tool that transforms text descriptions into visual art, offering users the ability to generate unique images from their ideas. 

📲SOCIAL MEDIA

🗞️MORE NEWS

• OpenAI rolled back a GPT-4o update that made ChatGPT overly flattering. The team is now refining training methods, gathering broader feedback, and adding user controls to make the model more honest, balanced, and customizable.

• Google’s Gemini chatbot now supports native image editing, letting users modify AI-generated or uploaded photos with tools like background changes and object replacements. The feature, rolling out globally, includes watermarking to reduce deepfake risks.

• Duolingo launched 148 new language courses using generative AI, more than doubling its offerings. What once took years now takes months, thanks to AI-driven content creation. The company is shifting toward an “AI-first” strategy.

• Xiaomi unveiled MiMo-7B, a 7B-parameter AI model built for math and coding tasks, outperforming some larger models. It uses advanced training methods, reinforcement learning, and long context handling, showing small models can rival big ones with smart design.

• Mark Zuckerberg plans to add a paid tier and ads to Meta’s new AI app, which now has nearly 1 billion users. Meta aims to grow engagement first, then monetize with premium features and product suggestions.

• Jack Krawczyk, former lead of Google’s Bard and Gemini, has joined Meta AI just as Meta launches its standalone AI app. He emphasizes building assistants with warmth, humility, and trust—not just information.