The race never ends. And 2025 illustrated how we entered into the competition between not just ordinary LLMs, but reasoning models and agentic systems. Agentic benchmarks, tool calling, and real-world execution moved to center stage. Open source surged to the forefront, with capabilities that often rival and sometimes surpass proprietary models. Chinese open models, in particular, are shaking up the market and adding fresh competitive energy across the field.

Another important direction of the year – world models. The community has finally mobilized the resources to build models that better understand the world. The hyper-focus on LLMs is finally starting to fade, making room for architectures grounded in dynamics, context, and action.

So let’s look at it one by one.

AI Concepts That Defined 2025

These are recaps of the key models and notable families of models from the first half of 2025.

Open source upgrade

1. Kimi K2, DeepSeek-R1, Qwen3 (+Coder), and GLM-4.5

Chinese companies – DeepSeek, MoonshotAI, Qwen, and Z.ai – kept the reasoning and agentic field open and accessible to everyone. They have turned everything into a battleground of the strongest agentic models ever. DeepSeek-R1 set the trend for open reasoning models, and then in the middle of summer other models followed:

Explore the full range of their capabilities in this article:

2. What is so special about Kimi K2 Thinking?

After four months of Kimi K2 release, in November 2025, Kimi K2 Thinking was issued – the newest and most capable version of Moonshot’s open-source thinking model, and one of the most versatile among Chinese open reasoners.

Here we trace Moonshot AI’s journey toward lossless long context and agentic intelligence, and explain what makes Kimi K2 Thinking a standout member of the Kimi family, including smart architectural solutions with precision and quantization twists.

3. Everything You Need to Know about GPT OSS

This year even OpenAI re-entered the arena of open source with GPT-OSS family of models released in partnership with Hugging Face. After a year dominated by Chinese open models, GPT-OSS marks OpenAI’s first serious open release since GPT-2. And it is already reshaping the ecosystem, raising big questions about strategy, standards, and the future of open reasoning models.

Here is what is under the hood and why it is so important globally. →

The rise of world models

1. What's New in World Models?

World models still remain a small but very-very promising field. Each new development shows how AI is learning to model the physical world and the logic of action. We’re tracking these breakthroughs to keep you ahead of the curve. This year, we saw some very interesting versions of world models:

2. What is PAN? How to Build a Better World Model?

However, what works best for building a world model remains unsolved, since this field is quite young. Researchers from Carnegie Mellon University, MBZUAI, and UC San Diego took a critical path through world modeling and proposed a new look at it, defining the main mission of world models as: “Simulating all actionable possibilities of the real world for purposeful reasoning and acting.”

They see that a better world model architecture should be: 1) hierarchical and multi-level, 2) mix continuous and discrete representations, and 3) be generative and self-supervised.

This approach, as the researchers argue, can lead to a PAN world model system – Physical, Agentic, and Nested.

3. What is LeJEPA? The Theory Upgrade JEPA Has Been Missing

When we talk about world models, we can’t forget Joint-Embedding Predictive Architecture, or JEPA, proposed by Yann LeCun as the centerpiece of his vision for building AI systems that can understand and reason about the world.

At the end of 2025, we finally got it – the full theory behind JEPA, straight from Yann LeCun’s latest article. Here, we explain in detail the rules for building a good JEPA, and how it turns into a verified, practical method: LeJEPA.

What are Guardian Models?

And finally, let’s talk about safety. Guardian models are often overlooked, but they’re responsible for defending today’s AI from malicious use, harmful or strange outputs, and hallucinations. They’re built into every serious AI deployment, and each takes a different approach to safety.

Here’s everything you need to know about guardian models, plus one that deserves special attention: DynaGuard, which can enforce any set of rules you give it at runtime.

Also, check out our previous recap of the key concepts and methods from 2025 for the full picture.