Model News Q1 25

This article summarizes notable developments and updates in AI sota foundation models during Q1 2025.

The first quarter brought enhanced reasoning, agentic capabilities, and multimodal integration. These improvements help building more intelligent, autonomous systems.

Major Releases

Google DeepMind: Gemini 2.5

• Released March 25, 2025. See Gemini 2.5 blog post
• As of then it leads the LMArena leaderboard
• Introduces “thinking models” capable of reasoning before responding

OpenAI: GPT 4.5, 4o

• Released GPT-4.5 as new flagship model (for premium users) on February 27, 2025
• Announced April 18, 2025. See GPT-4o image generation announcement
• Adds native diffusion-based image generation directly within GPT-4o's interface
• Supports image editing through natural language prompts and inline multi-modal context handling

Open Source: DeepSeek R1 and DeepSeek-V3

• Fully open model with performance comparable to OpenAI-o1
• It demonstrates complex reasoning capabilities and is released under the MIT License for broad use and commercialization
• More details: DeepSeek R1 announcement, Wikipedia, DeepSeek V3 announcement

---

Trends

• Agentic AI: Models improve in autonomously planning, useing tools, and perform multi-step workflows
• Multimodal Integration: Unified handling of text, images, audio, and video for richer interaction and analysis
• Efficiency: Smaller models like o4-mini deliver strong results with lower compute and faster inference