Liquid AI challenges the size race: LFM2's hybrid architecture aims to change the rules

For the past few years, the large language model industry has lived by a simple principle: more parameters, better results. OpenAI kept scaling GPT, Google kept inflating Gemini, Meta kept enlarging Llama*. But in 2026 this approach is increasingly running into physical limits: data center power consumption, memory costs, inference speed. Boston-based startup Liquid AI believes it has found a way out of the dead end, and its new model LFM2-24B-A2B is, in its view, not just another release but a bid for an architectural revolution.

Liquid AI grew out of MIT research on so-called "liquid neural networks," inspired by biological nervous systems. Unlike classic transformers, where each layer performs a fixed operation, liquid networks can adapt their computation depending on the input data. It is a fundamentally different approach to processing information, and the team has been developing it consistently for several years. LFM2-24B-A2B is the culmination of that work: a model with 24 billion parameters built on a hybrid architecture that combines the classic attention mechanism with convolution operations.

To see why this matters, it helps to understand the problem. Standard transformers, which underpin GPT, Claude and other models, use a self-attention mechanism that lets every token "look at" all the other tokens in the context. It is a powerful tool, but its computational complexity grows quadratically with context length. Double the context window and compute costs quadruple. That is why working with long documents remains one of the most resource-intensive tasks for modern LLMs. Convolutional layers, by contrast, process information locally and scale linearly, but have historically been considered less expressive for language.

Liquid AI's hybrid approach tries to take the best of both worlds. The convolutional components handle local patterns: syntactic structures, short dependencies, repeated templates. The attention mechanism kicks in where the model needs to catch long-range links in the text: references to previously mentioned entities, logical chains, complex reasoning. The "A2B" label in the model's name points to a specific configuration of this balance between attention and convolution blocks. In essence, the model decides for itself which type of processing to apply to a given piece of data, which makes computation considerably more efficient.

Twenty-four billion parameters is a relatively modest figure by 2026 standards, when flagship models operate with hundreds of billions and even trillions of parameters. But that is precisely Liquid AI's central thesis: architectural efficiency matters more than brute force. If a model with 24 billion parameters can compete with models several times its size at significantly lower inference costs, that changes the economics of the whole industry. Fewer GPUs to serve requests means lower API costs. Less power consumption makes it easier to deploy the model on edge devices. Faster inference means a better user experience.

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

Still, a healthy dose of skepticism is warranted. The full benchmarks for LFM2-24B-A2B have yet to be examined and independently reproduced. Architectural innovations often look impressive on paper but run into unexpected problems when scaled in production, from training complexity to compatibility with existing optimization infrastructure. The ecosystem of tools around transformers took years to build, and any alternative architecture will have to prove its viability outside the lab as well.

Nevertheless, the direction Liquid AI is setting seems inevitable. The AI industry has reached a point where adding parameters yields diminishing returns, while demands for energy efficiency and speed keep growing. Companies that find a way to do more with fewer resources will define the next chapter of AI. And hybrid architectures like LFM2 may well turn out to be the key to that future.

*Meta is designated an extremist organization and banned in Russia.