4 Outstanding Families of Models You Must Know About

When you explore a full family of models, you start to see the ideas and strategies of the companies behind them. 2025 brought many notable newcomers to the landscape. Since it's impossible to cover everything, in our AI 101 guides, we’ve narrowed our focus to four model families that showcase distinctive approaches to AI model design.

These include SmolLMs by Hugging Face, Qwen Models, Liquid Foundation Models, and the classic BERT. It’s time to revisit (or explore for the first time) the journey of these models and their creators! →

1. Inside the family of Smol models

A deep dive into Hugging Face’s Smol family of models, particularly into SmolLM2, uncovering how smart training and clever data choices power reasoning in small language models (SLMs) on par with larger ones.

2. What is Qwen-Agent framework? Inside the Qwen family

Here’s the entire journey of Qwen models toward strong reasoning, matching or even being better than state-of-the-art models from OpenAI and DeepSeek. Plus, we discuss something that you might have missed: the Qwen-Agent framework – a full‑fledged agentic ecosystem that lets Qwen models autonomously plan, call functions, and execute complex, multi‑step tasks right out of the box.

3. Can Liquid Models Beat Transformers? Meet Hyena Edge – the Newest Member of the LFM Family

New approaches can keep the whole AI field moving. So it’s really important to know about Liquid Foundation Models (LFMs) and Hyena by Liquid AI, as they offer promising alternative to Transformers, built from the first priciples.

Unlike attention-based models, which struggle with speed and memory at scale, LFMs use dynamic systems that process data more like continuous signals – stable, fast, and well-suited for long-form inputs like text, audio, or video. Hyena and its newer version, Hyena Edge, replace attention with fast convolutions and gating, allowing high-quality reasoning even on everyday devices like phones.

4. Decoding BERT: From Original NLP Game-Changer to Today's Efficient AI (feat. ConstBERT)

Also, don’t forget about the powerful classics like BERT, Bidirectional Encoder Representations from Transformers, which was the first to:

• Pre-train a deep Transformer in a truly bidirectional way to understand context more deeply (it processes tokens both from left to right and right to left);

• Bring the “pre-train then fine-tune” paradigm to Transformers at scale.

Recently, BERT has sparked a wave of novel models, like ModernBERT, NeoBERT, and ConstBERT.

Here, we dive into the core ideas behind BERT and the diversity of its variants.