5+ Notable Models of 2025 that You Should Know (AI 101 Guide Recap)

During this month, we’re refreshing AI topics that have been trends and centers of attention in the first half of 2025. Our previous recap was about AI methods and techniques (check it our here) and now we’re moving to the first part of models’ reviews.

Today, we invite you to focus on models, full ecosystems, and tools from the AI 101 guides that stood out in January–June 2025 thanks to their unique approaches. AlphaEvolve, Codex, Mixture-of-Mamba, World and Reasoning models – how do they work and what makes them special? Time to explore the spotlights!→

1. What is Cosmos World Foundation Model Platform?

Physical AI is a field many companies are moving into, and NVIDIA is one of them. At the beginning of 2025, they unveiled an entire ecosystem called the Cosmos World Foundation Model (WFM) platform, which continues to evolve. It includes tools and models for creating, training, and deploying WFMs for Physical AI. Explore how it all works here:

2. What are World Models?

Meanwhile, in this article, we uncover the main things you need to know to build world models. A "world model" is a learned simulation of an agent’s environment that an agent uses to “imagine” and explore possible sequences of actions. We dive into how these models work, using the examples of the most notable current world models: Google DeepMind’s DreamerV3 and Genie 2, NVIDIA Cosmos WFMs, and Meta’s Navigation World Model (NWM).

3. Reasoning Models - Just Advanced LLMs or New Species?

One of the main shifts of 2025 is moving AI to the level of Reasoning Language Models (RLMs) that generate answers slowly and more thoroughly, learn through trial and error, can self-reflect, and use external tools. Could they be defined as a separate category of models? We figure it out here, plus break down the list of current powerful RLMs.

4. What is Mixture-of-Mamba?

Mixture-of-Mamba (MoM) is a great example of a hybrid model. It expands Mamba architecture’s capabilities by incorporating Mixture-of-Transformers concept and modality-aware sparsity, making fast and efficient Mamba multimodal.

5. What's Cool About AlphaEvolve and Codex?

AlphaEvolve and Codex are more tools than models, and they also deserve your attention.

• AlphaEvolve is Google’s evolutionary coding agent designed to autonomously discover novel algorithms and scientific solutions. It can optimize engineering algorithms, like the building of hardware, and solve complex problems.

• Meanwhile, OpenAI’s Codex is the first cloud-hosted Autonomous Software Engineer (A-SWE). It’s like your virtual coworker within ChatGPT that works safely with your repository, writing, testing, and fixing code.

Here we dive into their unique workflow.