Liquid AI challenges the race for bigger models: LFM2's hybrid architecture changes the rules

For the past few years, the large language model industry has operated on a simple principle: more parameters equal better results. OpenAI has scaled up GPT, Google has expanded Gemini, Meta has increased Llama. But in 2026, this approach is increasingly running into physical constraints — data center energy consumption, memory costs, inference speed. Startup Liquid AI from Boston believes it has found a way out of this impasse, and its new LFM2-24B-A2B model is not just another release, but a claim to an architectural revolution.

Liquid AI is a company that grew out of MIT research in so-called "liquid neural networks," inspired by biological nervous systems. Unlike classical transformers, where each layer performs a fixed operation, liquid networks are capable of adapting their computations depending on input data. This is a fundamentally different approach to information processing, and the team has been consistently developing it for several years. LFM2-24B-A2B became the culmination of this work — a model with 24 billion parameters, built on a hybrid architecture that combines the classical attention mechanism with convolutional operations.

To understand why this matters, you need to grasp the problem. Standard transformers, which underlie GPT, Claude, and other models, use a self-attention mechanism that allows each token to "look at" all other tokens in the context. This is a powerful tool, but its computational complexity grows quadratically with context length. Double the context window — and you get a fourfold increase in computational costs. This is why working with long documents remains one of the most resource-intensive tasks for modern LLMs. Convolutional layers, on the other hand, process information locally and scale linearly, but have historically been considered less expressive for language tasks.

Liquid AI's hybrid approach tries to take the best of both worlds. Convolutional components handle the processing of local patterns — syntactic structures, short-range dependencies, recurring templates. The attention mechanism engages where it's needed to capture long-range connections in text — references to previously mentioned entities, logical chains, complex reasoning. The "A2B" designation in the model name points to a specific configuration of this balance between attention and convolution blocks. Essentially, the model itself decides what type of processing to apply to a particular fragment of data, making computations significantly more efficient.

Twenty-four billion parameters is a relatively modest number by 2026 standards, when flagship models operate with hundreds of billions and even trillions of parameters. But that's precisely Liquid AI's main thesis: architectural efficiency matters more than brute force. If a model with 24 billion parameters can compete with models several times larger while consuming significantly fewer inference costs, this changes the economics of the entire industry. Fewer GPUs to service requests — lower API costs. Less energy consumption — easier to deploy the model on edge devices. Faster inference — better user experience.

For the industry as a whole, the release of LFM2-24B-A2B fits into a broader trend. More and more research groups and companies are coming to the conclusion that the era of "dumb scaling" is ending. Mamba and other state-space-based architectures, work on sparse models with Mixture of Experts, quantization and distillation — all of these are attempts to extract more intelligence from fewer computations. Liquid AI is going its own way, and its hybrid approach looks like one of the most elegant solutions to the problem.

That said, it's worth maintaining healthy skepticism. Full benchmarks of LFM2-24B-A2B still need to be studied and independently reproduced. Architectural innovations often look impressive on paper but encounter unexpected problems when scaled into production — from training complexity to compatibility with existing optimization infrastructure. The ecosystem of tools around transformers was built over years, and any alternative architecture will need to prove its viability not just in the lab.

Nevertheless, the direction that Liquid AI is setting seems inevitable. The artificial intelligence industry has reached a point where adding more parameters yields diminishing returns, and demands for energy efficiency and speed only grow. Companies that find ways to do more with fewer resources will define the next chapter in AI development. And hybrid architectures like LFM2 could well prove to be the key to that future.